Imagine you are driving toward an intersection. A truck parked on the corner blocks your view of a car speeding toward the intersection from the right. You can’t see it, but the car next to you can. What if that neighboring vehicle could simply tell your car,

“Vehicle merging from the right, about 15 meters away, moving quickly.”

Your car now knows something its own sensors could never observe.

This simple idea of cars helping each other by describing what they see in natural language motivates UNCAP, our new framework for cooperative autonomous driving. But there’s an important twist: not every neighboring vehicle is worth listening to, and not every message deserves the same level of trust.

UNCAP teaches autonomous vehicles both who to talk to as well as how much to trust what they hear.

Why should cars talk with language instead of sharing sensor data?

Researchers have since long envisioned fleets of autonomous vehicles cooperating to make driving safer. Blind corners, occlusions, and hidden pedestrians will become much less dangerous if nearby cars can share what they observe.

The traditional approach is straightforward: cars can exchange raw sensor data such as camera images, LiDAR point clouds, or radar measurements. However, this quickly becomes impractical to be implemented at scale for a variety of reasons. Raw sensor streams consume enormous bandwidth, require significant computation to process stream data, and assume that every vehicle has comparable sensing hardware. In reality, with the rise of autonomous driving, the roads are instead observing a mixture of manufacturers, sensor configurations, and heterogenous software stacks. Natural language offers an intriguing alternative to sensor data sharing. Instead of transmitting megabytes of images, a vehicle might simply say, “pedestrian crossing behind bus” or “motorcycle approaching from the left, fast-moving”. These messages are tiny, hardware-agnostic, and immediately understandable by modern vision-language models across the spectrum.

But language introduces a new problem. At a busy intersection, a vehicle might have ten or twenty neighbors. Should it listen to all of them? And if two vehicles disagree, whose description should it believe? UNCAP answers all these questions!

How UNCAP works

An appeal of UNCAP is that it works entirely in a zero-shot manner. Instead of training a new model (which is resource intensive), UNCAP has a “Bring-your-own-model” strategy. It builds on existing vision-language models and adds a communication strategy that allows vehicles to cooperate intelligently. The UNCAP framework includes four stages.

UNCAP helps each vehicle decide which neighbors matter. Those messages describe nearby vehicles, motion, and confidence, which UNCAP combines before making a driving decision.

1. Discover nearby vehicles

Each vehicle periodically broadcasts a tiny “heartbeat” of information containing only its position and direction of travel. This heartbeat consists of only a few bytes of data. The aim is to provide just enough context for nearby vehicles to understand who is around them.

2. Decide who matters

On the road, most nearby vehicles are irrelevant. A car traveling away from the intersection provides little useful information, whereas a vehicle approaching the same intersection may offer exactly the observations we need. UNCAP reasons about which neighbors are both close enough and likely to affect the vehicle’s future path. This is crucial as it dramatically reduces the bandwidth of unnecessary communication as well as the computation required by downstream stages. It enables UNCAP to scale to handle large teams.

3. Exchange observations and reason about trust

Relevant vehicles then describe what they see using natural language, accompanied by calibrated confidence estimates. Calibration is desirable as it produces reliable uncertainty estimates. UNCAP uses conformal prediction, which is a lightweight calibration method that ensures confidence values reflect the actual reliability of the information being received. Every vehicle then reasons about whether an incoming message actually reduces its uncertainty about the scenario. To answer this, UNCAP measures the mutual information contributed by each incoming observation. Mutual information ensures that the vehicle can give greater weightage to messages that make the scene clearer, while uninformative or potentially misleading observations are discounted. Rather than blindly treating every neighbor homogenously, each vehicle builds an uncertainty-aware understanding of the environment.

4. Make a driving decision

Finally, the scene description generated by reasoning about all incoming messages is passed to an off-the-shelf vision-language model, which produces a driving decision such as merge, wait, or yield, along with a corresponding confidence score.

What did this research teach us?

We evaluated UNCAP in the popular CARLA driving simulator using the OPV2V benchmark, which is a widely used benchmark by autonomous driving researchers. OPV2V includes highway merging, urban intersections, and other scenarios where vehicles frequently experience occlusions.

These bird’s-eye-view (BEV) examples compare driving with and without UNCAP. With selected helper messages, vehicles behave more proactively in merging and intersection scenes.

Compared with non-communicating vehicles, and compared with prior language-based approaches that blindly broadcast every observation to every other car, UNCAP consistently improved both efficiency and safety. Over various scenarios, our experiments yielded key qualitative gains of UNCAP over traditional cooperation schemes include:

• 63% less communication bandwidth
• 31% higher driving score
• 61% lower decision uncertainty
• 4× larger safety margins during near-miss events

UNCAP avoids broadcasting every observation to every nearby vehicle. It sharply reduces bandwidth while preserving the information needed for a driving decision.

Perhaps the most surprising result was that sharing raw images actually performed worse than sharing language.

Modern vision-language models appear to struggle when jointly reasoning over multiple camera feeds captured from different viewpoints. Translating those observations into concise natural-language descriptions first seems to act as a form of structured reasoning, distilling the essential information before it reaches the decision-making model.

Equally encouraging, the improvements were consistent across several underlying vision-language models, including GPT-4o, GPT-4.1, GPT-4o-mini, and GPT-5. This suggests that the benefits stem from the communication strategy itself rather than from any particular model.

Looking ahead

UNCAP is an early step toward autonomous vehicles that cooperate more like human drivers.

Today’s implementation has important limitations. Our experiments are limited to simulation, so real-world constraints such as noisy sensing, unreliable wireless communication, network delays and packet losses remain to be explored. Finally, because querying today’s vision-language models takes roughly a second, decisions are made at key moments rather than continuously.

That being said, the broader message extends well beyond autonomous driving.

As AI systems increasingly operate in teams, whether fleets of robots, drones, or software agents, the challenge is no longer simply making each agent smarter. The critical need of the hour is to ensure that each agent in the team learns how to communicate efficiently, decide which information to trust, and build a shared understanding of the world.

In a nutshell, UNCAP suggests that natural language, paired with principled reasoning about uncertainty, can provide a surprisingly effective foundation for that future.

Read the paper in full

UNCAP: Uncertainty-Guided Neurosymbolic Planning Using Natural Language Communication for Cooperative Autonomous Vehicles, Neel P. Bhatt, Po-han Li, Kushagra Gupta, Rohan Siva, Daniel Milan, Alexander T. Hogue, Sandeep P. Chinchali, David Fridovich-Keil, Zhangyang Wang, Ufuk Topcu.

This work was nominated for the AAMAS 2026 best paper award.

By Wai-keung Fung, Cardiff Metropolitan University

At the recent World Humanoid Robot Games 2026 in Beijing, robots sprinted, boxed, played football and tackled tasks inspired by factories, restaurants and emergency situations. The robot Tiangong Ultra even completed the 100 metres in 8.64 seconds, quicker than Usain Bolt’s 9.58-second world record.

But if we really want to know whether humanoid robots are ready for everyday life, a far less glamorous challenge may tell us more: doing the laundry. Can a robot take a crumpled T-shirt from a laundry basket, work out which way round it is, fold it neatly and recover when the sleeve slips from its hand?

No robot system can yet reliably handle the whole laundry process, from carrying a basket to the washing machine and sorting clothes for different programmes to taking them out, hanging them to dry, folding and ironing them. Even folding remains a challenge: robots need to cope with clothes of different shapes, sizes, materials and levels of creasing.

Running fast is difficult, of course. A humanoid has to coordinate many joints, keep its balance and generate powerful movements without falling over. The task itself, however, is fairly clear. The robot knows where to go, the track is predictable and the finish line does not suddenly change shape.

Laundry is a very different problem. A shirt may be inside out. One sleeve may be hidden underneath another garment. A pair of trousers may be twisted around a towel. Fabric is also soft and constantly changing shape. Pick up one corner of a T-shirt and the rest of it moves.

Humans deal with this without giving it much thought. We shake a shirt out, notice where the sleeves are, feel when the fabric slips and adjust our grip almost automatically. For a robot, every one of those little actions is a problem to solve. This is why an everyday chore can reveal more about robot intelligence than a spectacular athletic feat.

A useful household robot cannot simply replay the same movement again and again. It has to notice when reality is different from what it expected and adapt. That’s important to consider because humanoid robots are increasingly discussed as future helpers in homes and care settings, as well as factories and warehouses.

A factory production line can often be arranged so that parts arrive in predictable positions. A family home is not like that: clothes, toys and furniture move around constantly, while people, children and pets make the environment even less predictable.

Why one robot cannot simply copy another

There is another difficulty. Once one robot learns how to fold a shirt, it might seem obvious that we should simply transfer that skill to every other robot. Unfortunately, physical skills do not work like an app that can be installed on any machine.

Imagine a specialist laundry robot with rigid arms, overhead cameras and grippers designed specifically for fabric. A humanoid may have cameras in its head, five-fingered hands and different joints and balance constraints. The same movement could therefore produce a very different result.

A fixed robot arm can reach forward without worrying about falling over. A humanoid may have to move its hips, legs and feet just to keep itself stable while reaching across a table.

Sensors are important too. One robot may have sensitive touch sensors in its fingers. Another may rely mainly on cameras.

Some knowledge can still transfer. One robot can teach another something about what a shirt looks like, where its shoulders are likely to be, what a good fold looks like or how to recognise when a sleeve is trapped underneath. The receiving robot still has to work out how to carry out those ideas with its own hands, joints, cameras and body.

Research projects such as Open X-Embodiment are exploring exactly this problem by pooling experience from many different robots. The project has brought together more than one million real robot trajectories from 22 different robot types. The hope is that robots can learn useful patterns from one another rather than starting from zero every time.

What should count as progress?

The Humanoid Robot Games are valuable because they make robotics visible and show how quickly the technology is moving.

We should simply be careful about what those demonstrations prove. For robots intended to live and work alongside us, better questions might be: can the robot deal with an unfamiliar object? Can it notice that it has made a mistake and recover without human intervention? How long can it keep working before someone has to step in?

These measures are unlikely to produce viral videos. They are, however, much closer to the abilities that will determine whether humanoid robots become genuinely useful.

A robot that can run 100 metres faster than the human world record time is impressive. A robot that can take a messy basket of laundry, sort it, fold it and recognise when things have gone wrong may tell us far more about whether general purpose robot intelligence has really arrived.

The most important race for humanoid robots may not be against the clock. It may be against the laundry pile.

Wai-keung Fung, Senior Lecturer in Electronics Robotics & Control Engineering, Cardiff Metropolitan University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Propeller-powered underwater vehicles have long helped scientists explore and monitor aquatic environments. But they’re limited by their own mechanics: spinning blades can snag on vegetation, stir up sediment, and startle the wildlife they’re often sent to study, making them poorly suited to shallow creeks, dense weeds, or close encounters with fish.

That’s one reason roboticists have spent years building machines that swim like fish instead, bending their bodies rather than spinning a propeller. The catch is that most fish-inspired robots are built for one size and one job, so scaling them up or down usually means starting from scratch.

A team of engineers at EPFL and New York University (NYU) says it has found a way to solve that problem. They’ve unveiled ScaFi (for Scalable Fish), a robot modeled on fish like cod and mackerel. These fish swim by concentrating most of their body bending toward the tail end, a style that, in nature, spans an unusually wide range of body sizes.

“Right now, if you want to monitor a creek and then monitor a lake, you basically need two different robots, built and tested from the ground up,” says Nana Obayashi, first author on the study published in npj Robotics. “The environments we care about don’t come in one size, so we don’t think the tools should either.”

Currently an assistant professor of mechanical and aerospace engineering at the NYU Center for Robotics and Embodied Intelligence, Obayashi led the project while completing her PhD in the Computational Robot Design & Fabrication Lab, led by Josie Hughes, in EPFL’s School of Engineering.

ScaFi at 0.6, 1.1, and 2.9 meters long. 2026 CREATE Lab EPFL. CC BY SA-4.0.

Studying different environments systematically

ScaFi has a rigid front section and a flexible tail made of fiberglass rods. A single motor pulls two tendons that cross near the tail’s end, producing the “S”-shaped bend required for fish-like swimming motion.

The diameter of the rods forming the tail are the only part that must change with the size of the robot. They grow proportionally thicker as the robot scales up, to preserve similar tail-bending behavior. The underlying motor mechanism and crossed-tendon system stay the same.

That matters because it could cut the engineering effort needed to build fish-like robots for different environments. It also gives researchers a platform for studying how swimming performance changes with scale, a question that’s hard to study systematically in animals or custom-made robots alike.

Into the water!

The team built three robots — roughly 0.6, 1.1, and 2.9 meters long — and tested how well each swam. In collaboration with EPFL’s Unsteady Flow Diagnostics Lab, led by Karen Mulleners, the researchers found that the smallest robot produced swirling water patterns similar to those left by real fish, and across all three sizes, swimming motion lined up closely once adjusted for body size. This suggests that that the scaling approach preserved the fish-like gait even as the robots grew nearly fivefold in length.

They also deployed the robots in the field: the medium-sized one in a Swiss stream, the largest on Lake Geneva, the smallest in creeks only 15–30 centimeters deep. During the stream test, the robot kept swimming even after a GPS dropout.

Energy efficiency proved harder to scale. The two smaller robots performed similarly, but the largest was consistently less efficient and needed a different, more powerful motor. The authors suggest drag and inertia may be to blame, though the exact cause is unresolved, meaning the team scaled the swimming motion itself more cleanly than the energy it takes to produce it.

A similar tradeoff showed up in disturbance tests. The smallest robot was most agile but recovered slowest after being knocked off course, while the larger robots were less nimble but more stable.

The same approach, scaling around one key structural parameter, could apply to other compliant robots, including ones outside water. Whether energetic performance can be scaled as successfully as the swimming motion remains an open question.

Funding

This research was partially funded by the European Union’s Horizon 2020 program under Marie Skłodowska-Curie grant agreement No. 945363.

Reference

Obayashi, N., Anastasiadis, A., Gumowski, J. et al. ScaFi: length-scalable, compliant, parametric robotic fish design for operation in multiple environmental niches. npj Robot (2026).

China is targeting the next industrial revolution by doubling down on embodied AI and humanoid robotics, building directly on its vast manufacturing infrastructure.

By Ellen H. Rumley and Allison Okamura

Beyond the Robot Olympics

Chinese robots have quickly entered global public consciousness in recent years, appearing as background robo-dancers in pop concerts and as humanoid marathon and sprint runners outpacing elite athletes (1). Domestically, robots are deployed as traffic guides and restaurant service staff, while swarms of drones entertain audiences in urban light shows or deliver takeout to high-rises (2, 3). These public debuts serve to both normalize automation in everyday civilian life (4) while projecting international soft power.

Behind the public spectacle lies a bold agenda for global economic dominance. In its geopolitical contest with the United States, Beijing sees robotics as the core engine to project power and dictate the future of global supply chains (5). This ambition is further accelerated by domestic drivers of an aging population and a shrinking industrial workforce, compounded by the legacy of the one-child policy and rapid urbanization. Alongside targeted talent acquisition initiatives like the K-Visa (6, 7), China aims to pioneer a new era of industrial leadership while mitigating a long-term structural gap through a state-orchestrated robotics surge.

Five-Year Plans (FYPs)

The Chinese Communist Party leverages its central authority to align state enterprise, private capital, and supply chains towards common strategic objectives, carrying the momentum to implement sweeping technological changes. Since the 1950s, every five years the government has produced large-scale roadmaps detailing major objectives mandated across its various levels of government. Called the Five-Year Plans (FYPs) – these documents provide bold economic development initiatives that are credited for structural reform like the opening of China to foreign trade in the 1980s (8). The FYPs have been influential in strengthening the nation’s manufacturing base and is now playing a defining role in advancing Chinese robotics.

FYP13 & FYP14: Made in China

The 13th FYP (2016-2020) laid the critical foundation of China’s robotics ascent and operationalized Made in China 2025 (MIC2025), a campaign to transition China from a “low-cost global factory floor” into a sovereign high-tech manufacturing leader (9). While most leading industrial nations historically outsourced production, China, as a historical receiver of outsourced production, took advantage of their vast manufacturing infrastructure to invest early into physical foundries and integrated supply chains that many nations are now spending heavily on to re-shore within their own borders (10). Key goals of MIC2025 were to climb up the global value chain across strategic sectors like robotics, using massive government subsidies to offset production costs (11). While many international critics condemned these subsidies as market distortion and imposed sanctions as a response (12), paradoxically this intensified Beijing’s efforts towards a sovereign tech sector.

Foreign frictions did ultimately lead Beijing to retire the branding “Made in China 2025”, but its core robotics objectives were carried out through focused sectoral plans in the 14th FYP (2021-2025). The Ministry of Industry and Information Technology (MIIT) even introduced two robot-specific initiatives: the “14th Five Year Plan for the Robotics Industry” (2021), which granted state-support for R&Ds in an effort to replace foreign suppliers of critical robotic components with domestic supplies, as well as the “Robot + Application Action Plan” (2023), which stimulated robotics demand by mandating adoption of domestic robotics across ten key sectors including agriculture, logistics, and health. This dual coordination of supply and demand, along with the push for general industrial leadership, enabled several impressive milestones to be achieved by 2025: China increased domestic market shares in industrial robots from 30% to 57% within a decade (13, 14), and domestic suppliers for industrial robots became market leaders for the first time (15). They reached 295,000 units of industrial robot installations in 2025 alone – 54% of all global installations that year – and entered the top five global manufacturers rankings for industrial robotics. Regional robotics hubs also flourished, including AI R&D clusters in Shanghai, hardware manufacturing bases in the Yangtze River Delta (16) and a growing startup ecosystem in Shenzhen (17) – creating a full-stack environment capable of rapidly prototyping fully integrated robotic systems. Local sourcing of many basic hardware components was further made possible by the massive mining and refining infrastructure that China possesses – which processes over 70% of the global production of rare earths like neodymium, which are critical for hardware like electromagnetic motors (18).

Bottlenecks in Precision Hardware & Semiconductors

The 13th and 14th FYPs propelled China into a position of industrial robotic leadership in production scale, installation quantity, and cost effectiveness. Despite their indisputable achievements, some critical milestones were not met. For one, domestic robotics supply chains did not approach the levels of self-sufficiency expected by 2025. While Chinese robotics command near totality of domestic market shares in lower-stakes industries like textiles, they captured closer to 30% of domestic market shares for higher stakes sectors like automotive industries (5), and even less so – at 15% – for high-end machine tools (19). This disparity stems from a historical deficit in high-precision engineering in China. Key components essential for durable industrial-grade robots, such as strain-wave reducers, precision gearboxes, and high-performance servo motors, remain reliant on imports from Japan and Germany (20), where precision manufacturing traditions are backed by decades of institutional standards and specialized labor (21, 22). The incoming 15th FYP does address this shortcoming, however, and prioritizes China’s advancement of precision hardware in the coming years (23).

Concurrently, Beijing fell short of its goal to achieve 70% self-sufficiency in semiconductor manufacturing, missing the mark by an estimated 40% (24). This discrepancy highlights a bottleneck for China’s robotics ambitions, which depend on advanced logic chips for model training. The sharp decline in Chinese purchases of U.S. AI chips by late 2025 reflects an accelerated push toward domestic chip sovereignty (25). Against the backdrop of export controls from the U.S. and rising tensions across the Taiwan Strait—the global leader of semiconductor manufacturing (26) – Beijing’s pursuit of computational autonomy serves to protect itself against supply chain chokepoints that could derail its incoming robotics plans.

FYP15: Embodied AI with a Human Form

The 15th Five Year Plan (2026-2030) marks a strategic shift from traditional advanced automation towards ‘embodied intelligence’ – defined as AI that physically interacts with real-world environments. Elevated to one of China’s top priority “future industries”, embodied AI is tasked with driving economic growth while modernizing national defense (27, 28). Elevating embodied AI from a niche subgoal to a central industry priority carries significant policy weight; central ministries, provincial governments, and state financial institutions are now mandated to tightly coordinate robotics development, reallocating substantial shares of state-backed venture capital and guidance funds to the sector (23).

Much like the 13th and 14th FYPs laid the groundwork for industrial automation, the 15th FYP launches a major campaign to accelerate physical AI through the “AI Plus Action Plan” (29). Under this initiative, Beijing is pushing for the widespread diffusion of low-cost, high-efficiency AI infrastructure across society – both domestically and across the Global South. AI integration is targeted across science, education, agriculture, green technology, defense, governance, and industrial development. Instead of pushing the frontiers of AI development, this policy prioritizes practical cross-sectoral applications that may already be feasible for domestic AI semiconductor chips – driving the iterative hardware advancements needed for Beijing’s long-term technology sovereignty (30). And as physical embodiment of AI, robots are expected to soon play a pivotal role in these directives (23). The real-world interaction of robots could provide vast physical dataset spillovers that further accelerate AI capabilities – a process supported by policies like “sandbox regulations” designating specific urban zones for real-world testing and training (31).

The 15th FYP specifically emphasizes the development of embodied AI in humanoid form factors. State subsidies are flooding the humanoid sector along large investments from tech giants like Huawei and Alibaba (32, 33), signaling a push from government and private entities alike towards the mass-deployment of humanoids in a manner reminiscent of the Chinese global domination of the electric vehicle (EV) market (34, 35). Owing also to the hardware surplus and supply chain synergies resulting from the EV push (36), China has already managed to scale its domestic humanoid sector to over 150 companies – resulting in 15,000 global installations, equivalent to 85% of global humanoid deployments in 2025 during the country’s first year of mass production (37, 38). Efforts for mass production were especially prominent at this year’s International Conference on Robotics and Automation (ICRA 2026), where Chinese humanoids and dexterous hand start-ups dominated the exhibit floor. Clearly, Beijing views humanoids as a leapfrog technology that could spearhead the next industrial revolution.

There are ongoing debates within the robotics community as to whether humanoids are the ideal form factor for general-purpose robots. On one hand, redundant joints for appearing human-like could introduce unnecessary mechanical inefficiencies and control complexity. On the other hand, humanoids could be, with enough data, an ideal form factor for interacting with everyday objects and urban spaces designed with people in mind.

Indeed, China’s focus on the humanoid form factor aligns closely with its data aggregation strategy: mapping directly to human workers allows a seamless transfer of human motion data into imitation learning models. In line, by July 2026, 40 humanoid training facilities had opened across major Chinese cities (39). In these hubs, humanoids from various domestic manufacturers are pooled to train basic skills – such as grasping and transport – in simulated real-world scenarios (40). Human operators wearing VR headsets and exoskeletons teleoperate the robots through task execution, capturing raw motion and spatial data. This supervised training is repeated hundreds of times daily, cleaned, labeled, and shared across manufacturers as foundational reference datasets. This shared public infrastructure enables scalable low-cost training useful for AI model developers – underpinning a broader national strategy towards leadership in generalizable physical AI (41).

Ultimately, Beijing is placing a large gamble on humanoids. While concentrating national resources has propelled China into the global market lead, it also exposes the sector to severe structural risks. Intense domestic competition can drive down hardware prices, squeeze profit margins and place firms at risk of overcapacity (42). Because policy-driven supply currently outpaces market demand, only a fraction of today’s 150+ companies are expected to survive the coming years (43, 44). While parallel R&D efforts can spur rapid innovation, the absence of a strong commercial demand risks investments from Beijing amounting to waste, which would further strain local government budgets.

This risk extends to investors as well: when government backing signals long-term viability before demand has caught up, the speculative value of humanoids can inflate beyond justification, creating conditions of volatility we see mostly recently playing out in the market. Unitree Robotics, the leading Chinese humanoid and quadruped manufacturer illustrates this pattern as it went public on the Shanghai Stock Exchange; shares surged as much as 629% intraday and has dropped by nearly 45% in the days following (45).

Ultimately, China’s 15th FYP represents a high-risk, high-reward strategy, leveraging state coordination and massive data scale to brute-force the commercial viability of humanoid robots. Whether Beijing succeeds in leapfrogging the world in embodied AI or stumbles under the weight of market overcapacity will carry high consequences for the international robotics industry and society at large.

Thanks for reading. Next up, we will be taking a road(map)-trip across the European Union and examine Brussels’ contrasting regulatory approaches to robotics.

Ellen H. Rumley – Policy Analyst

Allison Okamura – Vice President

IEEE RAS Science & Technology Watch Board

References

1. Ye, Ziyin. “Yearender: China’s Humanoid Robots Step from Spectacle toward Scalable Industrial Reality.” Xinhua, 31 Dec. 2025.
2. Che, Chang. “Inside China’s Robotics Revolution.” The Guardian, 19 Mar. 2026.
3. Yang, Zeyi. “Food Delivery by Drone Is Just Part of Daily Life in Shenzhen.” MIT Technology Review, 23 May 2023.
4. “How Different Countries Perceive Robots: An Original Survey.” Mitsubishi Research Institute, Feb. 2025.
5. Grant-Chapman, Hugh, et al. “Is China leading the Robotics Revolution?” ChinaPower Project, Center for Strategic and International Studies, 12 Feb. 2026.
6. “Robotics Revolution Driving Industrial Transformation.” China Daily, 10 Sept. 2025.
7. Bao, Anniek. “China’s K-Visa Plans Spark Worries of a Talent Flood.” CNBC, 1 Oct. 2025.
8. Yin, Ju and Xu, Jia. “Made in China 2025 and Industrial Upgrade.” SpringerLink, Springer, 2023.
9. Kuo, Kaiser. “Made in China 2.0: The future of global manufacturing?” World Economic Forum / J.H. Whitney & Co., 26 June 2025.
10. Amodio, Francesco, et al. “How China Became the World’s Factory: Trade, Industrial Policy, and Growth.” VoxDev, 18 Feb. 2026.
11. Xiang, Shuhui, et al. “Two Decades of Chinese Industrial Subsidies.” VoxEU, Centre for Economic Policy Research, 7 Jan. 2026.
12. “Made in China 2025 Dropped: Media Analysis.” MacroPolo, 15 Mar. 2019.
13. Tobe, Frank. “2014 Was Another Record Year for Robotics.” The Robot Report, 11 Apr. 2014.
14. Kim, Jenny, et al. “World Robotics R&D Programs.” International Federation of Robotics, 1 Jan. 2025.
15. “Top 5 Global Robotics Trends 2026.” International Federation of Robotics, 8 Jan. 2026.
16. Andreoni, Antonio, et al. “China’s Industrial Policy and Robotics Automation.” Global Media and Communication, SAGE Publications, 2 Aug. 2024.
17. Chang, Wendy, et al. “Embodied AI: China’s Ambitious Path to Transform Its Robotics Industry.” Mercator Institute for China Studies (MERICS), 30 Apr. 2026.
18. Gaudiaut, Tristan. “China’s Dominant Share in Global Mining Production of Critical Minerals.” Statista, 5 Nov. 2025.
19. “Was Made in China 2025 Successful?” Rhodium Group, 5 May 2025.
20. “Made in China 2025: Evaluating China’s Performance.” U.S.-China Economic and Security Review Commission, 14 Nov. 2025.
21. Tsuwa, Hideo, et al. “Precision Engineering and Industrial Automation Development.” Precision Engineering, ScienceDirect, Jan. 1979.
22. Lampe, Hannes W., et al. “Product Innovation and Industrial Policy in Emerging Markets.” Journal of Product Innovation Management, Wiley, 27 Nov. 2024.
23. Cheung, Sunny. “China’s New Five-Year Plan Prioritizes Robotics: The World Should Pay Attention.” The Diplomat, Mar. 2026.
24. Shivakumar, Sujai, et al. “China’s Localization Drive in Semiconductors Gains Impetus Allied Chip Export Controls.” Center for Strategic and International Studies, 24 Mar. 2026.
25. MacCarthy, Mark. “Ball Game’s Over: The US Is Out of the AI Chip Market in China.” Brookings Institution, 17 Jun. 2026.
26. Mickle, Tripp. “Taiwan, China, Chips, and Silicon Valley.” The New York Times, 24 Feb. 2026.
27. “China Makes AI-Powered Robots Core of National Strategy.” International Federation of Robotics, 5 May 2026.
28. Cheung, Sunny. “New Military-Civil Fusion Body for PRC Robotics Ecosystem.” China Brief, Jamestown Foundation, 18 Dec. 2025.
29. “State Council Policy Briefing on Technological Advancement and Industry Guidelines.” The State Council of the People’s Republic of China, 21 Aug. 2025.
30. Martin, Nik. “China Chips and Semiconductor Industry: US Technology Constraints and Artificial Intelligence.” Deutsche Welle, 20 Apr. 2026.
31. “China’s 15th Five Year Plan Explorer.” China FYP.
32. “Huawei Backs GigaAI in Latest Bet on China’s Booming Humanoid Robot Sector.” China Biz Insider, 28 Nov. 2025.
33. Cheng, Evelyn. “Alibaba Leads $100 Million Investment in Chinese Humanoid Robot Startup.” CNBC, 8 Sept. 2025.
34. Zhang, Tong, et al. “Empirical Analysis of Technological Innovation and Robotics Integration.” Journal of Asian Economics, ScienceDirect.
35. Lu, Fengming and Ma, Xiao. “Rise of China’s EV Industry.” The China Journal, No. 96, Jul. 2026.
36. Tobin, Meaghan and Keith Bradsher, Keith. “Why It’s Nearly Impossible to Build a Robot Without China.” The New York Times, 11 Jun. 2026.
37. Ye, Alice. “China’s Humanoid Robot Dominance: What It Means for the Global Economy.” China Biz Insider, 20 May 2026.
38. Tao, Ni. “Humanoid Robots in China: Progress, Limits, and Reality.” Cheung Kong Graduate School of Business, 23 Jan. 2026.
39. Zhou, Viola and Lo, Kinling. “Inside China’s Robot Training Centers for Workers.” Rest of World, 7 Jan. 2026.
40. Yiyi, Zhang. “China Ramps Up Embodied AI and Next-Gen Robotics Ecosystem.” Global Times, 27 Jul. 2026.
41. Hannas, William C., et al. China’s Embodied AI Development and Strategic Roadmaps. GovWhitePapers, 1 Dec. 2025.
42. Tobin, Meaghan and Wu, Xinyun. “Unitree and UBTech Lead China’s Robotics Expansion.” The New York Times, 17 Dec. 2025.
43. Chen, Laurie, et al. “China’s Humanoid Robots Aren’t Smart Enough to Take Your Job Yet.” Reuters, 27 Aug. 2026.
44. Jonas, Adam et al. “The Humanoid 100: Mapping the Humanoid Robot Value Chain.” Morgan Stanley Research, 6 Feb. 2025.
45. “China Robot Maker Unitree’s Post-Listing Slump Sparks Bubble Fears.” Reuters, 24 Aug. 2026.

By Maria Jose Galvez Trigo, Cardiff University and Paul Willis, Cardiff University

New BBC sitcom Ann Droid imagines a near future in which robots provide care and companionship to older people at home.

The series centres on Sue, a recent widow played by Sue Johnston, her hapless son Michael and Linda, an assistive care robot played by Diane Morgan. Linda is designed to provide daily support and companionship, with predictably comic results.

As researchers in social care, ageing and human-centred robotics, we obviously watched Ann Droid with a keen eye. Beneath the jokes are questions that researchers are already tackling.

Could robots help address loneliness among older people? Can a machine provide companionship without replacing human connection? While the series exaggerates what robots can currently do, some of the technology it depicts is already being tested.

The BBC’s Ann Droid sitcom.

A social enterprise in south-west England has been piloting Comfort Companions in partnership with Age UK South Gloucestershire. AI-generated personas offer conversation and guidance to older people living alone and at risk of loneliness.

It runs alongside Age UK’s volunteer programme, with befrienders helping older people learn how to use the app. The AI companions are intended to supplement the waiting list for human befrienders rather than replacing them.

Physical robots are being tested too. West Berkshire Council has been running a trial using robotic pets in its care homes.

Neither scheme offers robots who look like Linda. Current systems tend to be AI avatars on a screen or robotic cats or dogs, or speakers that provide a voice in the room. They do much less than Linda too. They can’t help someone up after a fall, wash them or get them dressed.

What technology does currently do best is monitoring and prompting. Systems that can detect a fall and alert another person are far more mature, while technology can also encourage someone to contact friends, make a phone call, or get out of the house.

Can a machine make you less lonely?

It’s worth considering this distinction because loneliness is not simply the same as being alone. It’s a personal feeling that our relationships are insufficient, accompanied by a desire for more or better social contact.

A robot companion, therefore, is unlikely to solve loneliness simply by being present. What matters is the quality of the interaction and whether technology can help someone maintain relationships with the people who matter to them.

This is one of the more interesting ideas in Ann Droid. Linda can sometimes strengthen Sue’s existing relationships rather than replace them, encouraging her to meet friends and plan activities outside the home. That’s potentially a more useful way to think about companion technology. The goal need not be to create an artificial friend who substitutes for a human one, but to help people remain connected to others.

The barriers are still substantial. Battery life is a limitation. Most humanoid robots manage between 90 minutes and five hours per charge when new. This is why an overnight camping trip featured in the sitcom is a fair test on the fantasy.

Robots are much better at some tasks than others. A machine may be able to play chess, for example, but picking up an unfamiliar household object in the real world remains surprisingly difficult.

The social barriers are even harder. Robots struggle with the social subtleties that people take for granted. Systems that learn by copying human behaviour can reproduce an action without understanding its intention. They also cannot reliably interpret facial expressions.

A review of systems designed to detect emotion from faces found that they performed worst when interpreting older faces, with anger and neutral expressions among the least reliably identified. This is important in care because recognising how someone is feeling can be crucial.

Would we trust a robot carer?

Ann Droid also plays on a more instinctive discomfort: should we trust a machine that’s designed to look after us? The sitcom playfully touches on public worries about interacting with robotic companions, including people’s distrust and worries about potential harms. For example, in one episode two robots enjoy a joke together about killing their human companions.

Research suggests that older people can see value in companion robots. But acceptance depends partly on whether people feel they remain in control. And appearance is a factor too.

In 1970, the roboticist Masahiro Mori proposed the idea of the “uncanny valley”. As machines become more human-like, we initially respond more positively to them. But when they look almost human without quite getting there, that warmth can turn into discomfort.

Ann Droid – dinner with the robot-in-law.

Robots have some obvious advantages. They don’t get bored, tired or impatient. They can handle dangerous tasks and they’re available at 3am. But they can’t provide something fundamental to care: a relationship in which both people choose to be there.

That’s the bond that no machine can supply. Robots may have a useful role in social care, but they’re worth having only when people want them, alongside human care and companionship. Otherwise, we risk finding an expensive technological answer to a much harder question: why do some older people have so little human contact in the first place?

Maria Jose Galvez Trigo, Senior Lecturer (Associate Professor) in Human-Centred Robotics and AI, Cardiff University and Paul Willis, Professor of Social Care, Cardiff University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

A recent ICRA panel titled “Surviving the Paper Deluge” brought together leading robotics researchers who have grappled with the overwhelming number of robotics papers being published today.

The discussion ranged from hard numbers on publication growth, through the promises and risks of large language models (LLMs), to radical proposals for reshaping peer review as we know it.

Panel chair Aude Billard pointed to rapid growth across major IEEE Robotics and Automation Society venues, with a roughly exponential curve beginning around 2017. An estimate for 2025 suggests around 70,000 papers containing the word “robotics.”

Billard noted that while in some fields, extreme specialization may be an acceptable survival strategy, robotics is inherently different. A robot is an integration of perception, control, manipulation, learning, hardware, interaction, safety, and deployment. If researchers can only stay current within narrow silos, the field risks losing one of its core strengths: the ability to connect ideas across domains.

Kunpeng Yao presented a case study funded by IEEE RAS Science and Technology Watch Board. The team tried to do something that borders on heroic in today’s publishing environment; that is, reading an entire subfield carefully over a single year.

Focusing on papers related to learning from demonstration, the team searched IEEE Xplore for relevant 2024 papers, screened the results manually, and identified 347 relevant papers. Of those, only 69, or about 20 percent, were judged to have made notable contributions.

According to the criteria used, ‘notable’ papers tended to offer new formulations, mechanisms, or theoretical guarantees; serious comparisons against strong state of the art; clear gains in robustness, transfer, or failure recovery; new teaching or data collection modalities; and convincing real-robot validation. Papers were less compelling when they simply relabelled existing approaches or failed to demonstrate their claims.

Yao summarized one lesson neatly: notability should be judged against the state of the art, not against “the newest vocabulary.”

That distinction is of particular importance in a field where new labels can travel quickly and a paper may sound fresh without actually moving the frontier. Conversely, useful work is sometimes hidden in less fashionable venues or written by less visible groups.

Degrees of hallucination

One question is whether large language models can help with literature review. Nadia Figueroa explained that LLMs can improve the mechanics of literature review. That is, they can help with search, retrieval, clustering, tables, summaries, and rough conceptual maps. Work that once took weeks or months can now sometimes be done in hours. For a new PhD student entering a field, that is a genuine reduction in the barrier to entry.

But Figueroa also identified hallucination as a major issue. First-order hallucination involves fake or incorrect references. More subtle is second-order hallucination, where the reference is real but the model misstates what the paper actually did. Also dangerous is third-order hallucination, where the LLM invents plausible but false commonalities across papers.

“The danger currently is not fake citations,” observed Figueroa. “It’s fake understanding.”

An LLM-based literature review can contain real papers and still misrepresent the field. It can be fluent, structured, and wrong. If researchers outsource not just the mechanics of reading but the cognitive act of comparison, judgment, and doubt, the field may produce more text while developing less understanding.

Salami and sandcastles

Meanwhile, Greg Dudek described the familiar problem of “salami slicing,” where a larger body of work is divided into “the thinnest possible slices” and spread across workshops, conferences, and journals.

For Dudek, the system rewards this behaviour. Students need papers to graduate. Early-career researchers need papers for jobs. Committees often face too much material to read deeply, so titles, venues, counts, and indices become tempting shortcuts.

But the result is costly for readers. Multiple papers repeat the same background, divide one contribution into fragments, and make it harder to reconstruct the full story.

His preferred remedy is simple in principle but difficult in practice: publish fewer, more integrated papers.

Dudek is under no illusion about how hard this will be. “There is no fix,” he said, if by fix we mean a return to a quieter world. Incremental reforms may help, but the flood is still rising. He compared small procedural fixes to building better walls around a sandcastle while “there’s a tsunami coming from the back of the room.”

Re-designing peer review?

Renaud Detry argued that the growth in publications partly reflects the growing presence of applied systems work in academic venues. That is not necessarily bad. Robotics advances through real systems, data, benchmarks, platforms, and the last-mile effort needed to make ideas work outside idealized settings.

The problem is that the same evaluation machinery often judges very different contribution types. A new algorithm, a carefully engineered system, a benchmark, a dataset, and an industrially relevant validation study should not all have to pretend to be the same kind of paper. Robotics may need clearer tracks for fundamental science, applications, infrastructure, benchmarks, and technical correctness.

Visibility is not value

Dongheui Lee addressed another vital part of the new publication ecosystem: arXiv, open-source releases, project pages, videos, blogs, and social media. These tools can broaden access, improve reproducibility, and help readers decide what to examine closely. But they also distort attention. Visibility can reflect institutional prestige, networks, speed, and self-promotion as much as scientific value.

Lee’s suggestion was pragmatic: keep expert peer review as a quality filter but use its signals better. Editorial boards and media teams could do more to promote strong papers that receive excellent reviews but do not come from famous labs.

Beyond the gatekeeper model

The most radical question came from Shigeki Sugano, who asked “What if we were to abandon peer review altogether?” In Sugano’s model, legitimate robotics and AI manuscripts would be uploaded first to an open archive, along with videos, code, and data.

Community evaluation, under verified identities, would happen in public. Journals and conferences would then certify high-value work rather than deciding what gets to exist.

“The question is not whether peer review is valuable,” he argued. “The question is whether it must remain the only gate to visibility.”

Not everyone agreed. Popularity bias, gaming, reputation effects, and low-quality papers flooding the system were highlighted by other panelists as possible risks. Still, Sugano’s radical proposal addressed real challenges around the ability of traditional accept-or-reject peer review models to scale to meet the deluge of papers flooding the robotics community’s information space.

No one on the panel pretended there was a clean fix. Instead, it emerged that the paper deluge is a tangle of volume, incentives, tools, evaluation, visibility, and culture.

Nevertheless, several useful principles emerged. LLMs may provide some mechanical support but carry risks of intellectual outsourcing. Read selectively and deeply. Reward quality before publication counts. Recognize different kinds of value in robotics research. Treat visibility as a possible indicator of worth, not as a verdict.

The open question

One major question remains: what happens to robotics research if papers continue to be published at the current rate?

There are clear benefits. More papers, more preprints, and more routes into publication can lower barriers to entry, help new researchers find a foothold, and make the field more open to a wider range of voices. Platforms such as arXiv have also made it easier for work to circulate before, or outside of, traditional publication channels.

But the risks are equally clear. Robotics is an integrative field, bringing together perception, control, manipulation, locomotion, hardware, safety, learning, and human-robot interaction. If researchers can only keep up with narrow slices of that landscape, the field may become more siloed. One result could be a growing reliance on off-the-shelf tools, including black-box commercial systems, without a deep understanding of the foundations on which those tools depend.

There is also the problem of memory. At a certain scale, no individual researcher can read far enough back, or broadly enough across adjacent domains. That increases the risk of duplicated work, overlooked insights, and important papers disappearing beneath the next wave of more fashionable topics.

This article originally appeared on IEEE RAS.

Florent Delgrange won the Best Blue Sky Paper Award at AAMAS 2026 for his work Foundation World Models for Agents that Learn, Verify, and Adapt Reliably Beyond Static Environments. We caught up with him to find out more about his vision for agent learning.

What is the topic of your Blue Sky Ideas paper and why is it an interesting area for study?

My Blue Sky Ideas paper asks a simple but difficult question: how can an autonomous agent keep learning as its world changes without quietly losing the guarantees that made its behavior trustworthy?

Reinforcement learning and formal methods address complementary parts of this problem. Reinforcement learning allows an agent to learn by trial and error and can scale to environments for which we could never write down every rule. However, the agent is usually asked to maximize a reward. A poorly specified reward can be exploited, and a high reward does not by itself tell us that a safety or coordination requirement has been satisfied. Reactive synthesis starts from the other end: given a model of the environment and a logical description of the intended behavior, it can construct a policy that is correct by design. The difficulty is that it normally needs an explicit, fixed model, which is precisely what an agent lacks in an open and changing world.

The paper proposes a research agenda that brings these traditions into one loop. A foundation world model would be learned from experience, but structured so that a verifier can reason about it. As the agent learns a policy, it would update its model, measure the reliability of its abstraction, and check whether the policy still satisfies its specification. The verifier’s feedback could reject an unsafe update, request data from an uncertain region, or trigger a revision of the model.

This matters because real environments do not politely remain as they were during training. Goals evolve, conditions change, and, in a multi-agent system, every adapting agent changes the environment perceived by the others. Reliability therefore cannot be a certificate obtained once at deployment. It has to be maintained as the agent continues to learn.

What is your vision for foundation world models?

To me, foundation refers first to reuse, not simply to size. I do not envision a larger video predictor trained on more trajectories. I envision a persistent and structured model that an agent can carry across tasks, policies, and changing populations of agents.

One way to think about it is as a map that records more than roads. It should also tell the agent which areas have been surveyed, which conclusions depend on those areas, and when a change in the world has made an old route unreliable. A useful foundation world model should do the same for decision-making: predict what may happen, expose the structure needed for reasoning, and quantify when those predictions are trustworthy enough to support a guarantee.

I’d want such a model to have three key properties. First, it should be calibrated: every learned abstraction should come with a measure of its error or coverage, so that the agent knows where formal conclusions remain valid. Second, it should be compositional: verified local dynamics, behaviors, and certificates should be reusable when a new task is assembled. Third, it should be semantically queryable: a formal requirement or high-level instruction should help the agent derive a suitable reward model, task-specific abstraction, or policy prior with little additional experience.

The durable idea is that the world model becomes a common substrate for learning, planning, and verification. It should help an agent act, but also identify the limits of its competence, gather evidence where those limits matter, and explain why a particular behavior can or cannot currently be certified.

How does this vision differ from current models?

Most current model-based approaches focus on learning for a particular task, environment, policy, or training distribution. Their main model learning objective is predictive accuracy: reconstruct an observation, forecast the next (latent) state, or generate imagined trajectories for planning. Those are important capabilities, but average predictive metric is not the same as fitness for a particular guarantee.

Consider a model for a scenario involving an agent interacting in a warehouse, where the world model predicts almost every transition correctly but misses a rare, dangerous interaction with a forklift. Its average error may be excellent while that transition is exactly the one that determines whether a collision-avoidance claim is valid. The reverse is also possible: a compact model may ignore colors and textures yet preserve everything needed to reason about routes and collisions. For verification, the relevant question is therefore not only “How accurate is the model?” but “Which conclusions does this accuracy justify, for this policy and this requirement?”

Foundation world models would make that connection an explicit design objective. Their abstractions would carry reliability information tied to the behavior being analyzed, and this information would be revised online as the data distribution changes. Previously verified components could be reused and composed, while specifications expressed in logic or language could guide which representation and policy the agent needs for a new task.

The main difference is therefore a change in role. Current models are primarily prediction tools. The model I envisage is a persistent, analyzable basis for learning, adaptation, and formal reasoning, with language models providing a complementary semantic interface. They could translate high-level instructions into candidate specifications or propose structured model updates, while the world model grounds these proposals in experience and the verifier checks their validity. Foundation world models may benefit from scale through broader task and environment coverage, but scale alone does not provide the structure or certificates required for reliability.

Could you give an example of how such a model might work?

Consider a delivery robot in a busy warehouse. Its task could be stated as: “eventually deliver the package while always avoiding collisions.” Instead of manually combining many bonuses and penalties, the system would translate that requirement into a reward model. Learning and verification would then start from the same description of the intended behavior.

As the robot receives observations, it would learn a compact representation of the warehouse and a policy that acts on that representation. A verifier would ask two related questions. Does the policy satisfy the delivery and collision-avoidance requirement in the learned world model? And is the learned abstraction accurate enough, along the routes that matter, for that conclusion to be trusted? A certificate is meaningful only when both answers are supported.

Now suppose a new forklift begins using a shortcut that was rarely visited during training, cluttering the way. Rather than treating an old prediction as a guarantee, the model should lower its confidence in that region. The verifier would detect this, withdraw the affected certificate, reject a risky policy update, direct exploration toward the shortcut, and reinstate a guarantee only after the world model has been recalibrated.

The paper also considers a more ambitious test-time loop. A language model could propose one or more small formal program describing the new dynamics. A model checker would test them, learn to compose with them, and return a counterexample or structural inconsistency when they are wrong. The language model could revise its hypothesis, the robot could collect targeted experience, and the cycle would repeat. In this division of labor, the language model proposes and the formal verifier checks.

We already have pieces of the loop, including formal reward translations, verifiable abstractions, safe policy-improvement methods, and program generation. Building an efficient end-to-end agent that keeps all of these pieces calibrated while it learns remains the research challenge.

What do you think the impact of such a framework on multi-agent systems could be?

Multi-agent systems make the problem both more urgent and more difficult. Each learning agent is a moving part of every other agent’s environment. Even when the physical world is unchanged, the effective dynamics evolve as agents update their policies, join or leave the system, share information, or pursue new objectives.

A foundation world model could retain reusable descriptions of these interaction patterns, while formal specifications state what must hold for the group. A fleet of warehouse robots, for example, may need to avoid collisions and complete deliveries while respecting shared capacity constraints. The model could connect each local policy to the assumptions on which its certificate depends. If one robot changes its route, the system could identify which assumptions and guarantees are affected, collect new data where needed, and revise only the relevant components rather than relearning and re-verifying the entire fleet.

Composition is the main source of potential leverage. Verified local dynamics or coordination behaviors could serve as building blocks for new teams and tasks, and previous agents could provide useful priors for new participants. This could support faster adaptation while making failures easier to diagnose: the system should report which interaction invalidated a certificate and provide a counterexample, rather than only revealing that the joint reward has fallen.

The central obstacle for multi-agent systems is scalability. The joint state space grows very quickly with the number of agents, and local guarantees do not automatically compose into a global one. We will need principled ways to expose dependencies, preserve soundness under composition, and run verification quickly enough to influence learning online. If we can solve those problems, the field could move from learning coordination strategies and checking them afterwards to learning new strategies while continuously tracking which global properties remain guaranteed.

About Florent

Florent Delgrange is a postdoctoral researcher in computer science at the Artificial Intelligence Lab of Vrije Universiteit Brussel (VUB). His research lies at the intersection of reinforcement learning, world models, and formal verification. He develops methods for agents that can learn and adapt while justifying and certifying the behavior they adopt. He completed a joint PhD at VUB and the University of Antwerp in 2024 on the formal verification of deep reinforcement learning policies. His paper Foundation World Models for Agents that Learn, Verify, and Adapt Reliably Beyond Static Environments received the Best Blue Sky Paper Award at AAMAS 2026.

AURA Foresight has been selected as a finalist in the prestigious XPRIZE Wildfire Autonomous Wildfire Response competition, emerging as one of just four teams remaining from more than 130 teams from around the world.

XPRIZE Wildfire is a four-year, US$11 million global competition designed to accelerate breakthrough technologies capable of ending destructive wildfires. The Autonomous Wildfire Response track, worth US$5 million, challenges teams to autonomously detect, verify and respond to wildfire ignitions across a 1,000 km² landscape within just ten minutes. The finals will take place in Nenana, Alaska, where teams will demonstrate their technologies in realistic wildfire response scenarios.

Being selected as a finalist places AURA Foresight among a small group of innovators developing the next generation of tools to help communities, land managers and firefighters respond to wildfires faster and more effectively.

As wildfires become more frequent and severe worldwide, AURA Foresight is developing technology that helps stop fires before they become disasters. The team’s autonomous wildfire intelligence and intervention system is designed to detect ignitions within minutes and coordinate a rapid response while fires are still small and manageable.

Detect Early. Verify Quickly. Intervene Immediately.

AURA Foresight combines fixed sensors, artificial intelligence and swarms of autonomous flying robots into a seamless detect–verify–act capability.

Rather than fighting large wildfires after they have taken hold, the system is designed to identify potential ignitions, verify threats and intervene at the earliest possible stage, when a small spark can still be safely contained before becoming a major incident.

The platform continuously monitors landscapes using optical and thermal sensors, automatically identifies potential ignitions using AI, dispatches flying robots to verify threats, and coordinates rapid intervention before small fires can grow into destructive wildfires.

Unlike many wildfire technologies that rely on specialised infrastructure or proprietary hardware, AURA Foresight has been designed from the outset to be practical, affordable and easy to deploy.

Key advantages include:

• Out-of-the-box deployment with minimal setup requirements.
• Adaptability to any environment, from forests and remote wilderness to critical infrastructure corridors.
• Scalability from a handful to dozens of flying robots, depending on operational needs.
• Compatibility with different aircraft and robot platforms, avoiding vendor lock-in.
• Built on commercially available technologies, making advanced wildfire protection accessible and achievable for organisations of all sizes.

The result is a system that can help agencies move from passive monitoring and delayed response to proactive, automated intervention.

Developed With Firefighters, For Firefighters

AURA Foresight’s approach is shaped by close collaboration with firefighters and emergency response professionals. The consortium works directly with Lancashire Fire and Rescue Service in the United Kingdom and with wildfire response experts in Canada and Australia to ensure the technology addresses real operational challenges and integrates effectively into existing emergency response workflows.

This combination of frontline expertise and advanced robotics helps ensure that the technology remains focused on supporting firefighters and enhancing their capabilities in the field.

“It’s amazing what can be achieved when great people come together around a shared purpose,” said Dr Georgios Tzoumas, Team Co-Lead of AURA Foresight and Research Fellow at the University of Bristol.

“Our team has been working on autonomous wildfire detection and intervention technologies for more than six years. Reaching the XPRIZE finals is an exciting milestone because it feels like the solution is finally within reach. By combining AI, swarm robotics and close collaboration with firefighters, we’re showing that it’s possible to tackle wildfires while they’re still small, before they grow out of control. We can’t wait to demonstrate our technology in the field in Alaska.”

A Global Consortium Combining Expertise from the UK and Australia

AURA Foresight is a unique international collaboration bringing together AURA in the United Kingdom and Fire Foresight in Australia.

The consortium includes the University of Bristol, Bristol Robotics Laboratory, the University of Sheffield, Lancashire Fire and Rescue Service, Fire Foresight, SkyFly Drones, Southern Denmark University, Manchester University, Indicium Dynamics, Robotic Cats, Taz Drone Solutions and Little Place Labs.

Fire Foresight CEO Rob Vernon commented, “Our approach with the XPRIZE competition has been to bring all the pieces of the wildfire jigsaw puzzle together, and that capability only truly scales when it’s done with partners from around the world and on the ground. In Australia, we have been building the largest network of bushfire detection capabilities and the partnerships that we’ve formed, and continue to form, through this competition allow us to focus our attention on the autonomous mitigation and suppression capabilities, so desperately needed to address the challenges faced by wildfire and climate change”.

Together, these partners combine world-leading expertise in swarm robotics, autonomous systems, artificial intelligence, computer vision, aerial robotics, wildfire operations and field deployment.

By bringing together researchers, engineers, firefighters and technology innovators from across two continents, AURA Foresight is demonstrating how international collaboration can accelerate solutions to one of the world’s most pressing climate challenges.

As destructive wildfires continue to increase in frequency and severity around the world, AURA Foresight’s progression to the XPRIZE Wildfire finals highlights a new approach to wildfire management: tackling fires in their earliest moments, before they can grow out of control and threaten communities, ecosystems and critical infrastructure.

Our motto is – “Detect early. Verify quickly. Intervene immediately. Stop wildfires before they become disasters.” Fingers crossed for a good performance in our finals this weekend!

By Michael Psenka, Mike Rabbat, Aditi Krishnapriyan, Yann LeCun, Amir Bar

GRASP is a new gradient-based planner for learned dynamics (a “world model”) that makes long-horizon planning practical by (1) lifting the trajectory into virtual states so optimization is parallel across time, (2) adding stochasticity directly to the state iterates for exploration, and (3) reshaping gradients so actions get clean signals while we avoid brittle “state-input” gradients through high-dimensional vision models.

Large, learned world models are becoming increasingly capable. They can predict long sequences of future observations in high-dimensional visual spaces and generalize across tasks in ways that were difficult to imagine a few years ago. As these models scale, they start to look less like task-specific predictors and more like general-purpose simulators.

But having a powerful predictive model is not the same as being able to use it effectively for control/learning/planning. In practice, long-horizon planning with modern world models remains fragile: optimization becomes ill-conditioned, non-greedy structure creates bad local minima, and high-dimensional latent spaces introduce subtle failure modes.

In this blog post, I describe the problems that motivated this project and our approach to address them: why planning with modern world models can be surprisingly fragile, why long horizons are the real stress test, and what we changed to make gradient-based planning much more robust.

This blog post discusses work done with Mike Rabbat, Aditi Krishnapriyan, Yann LeCun, and Amir Bar (* denotes equal advisorship), where we propose GRASP.

What is a world model?

These days, the term “world model” is quite overloaded, and depending on the context can either mean an explicit dynamics model or some implicit, reliable internal state that a generative model relies on (e.g. when an LLM generates chess moves, whether there is some internal representation of the board). We give our loose working definition below.

Suppose you take actions and observe states (images, latent vectors, proprioception). A world model is a learned model that, given the current state and a sequence of future actions, predicts what will happen next. Formally, it defines a predictive distribution on a sequence of observed states and current action :

that approximates the environment’s true conditional . For this blog post, we’ll assume a Markovian model for simplicity (all results here can be extended to the more general case), and when the model is deterministic it reduces to a map over states:

In practice the state is often a learned latent representation (e.g., encoded from pixels), so the model operates in a (theoretically) compact, differentiable space. The key point is that a world model gives you a differentiable simulator; you can roll it forward under hypothetical action sequences and backpropagate through the predictions.

Planning: choosing actions by optimizing through the model

Given a start and a goal , the simplest planner chooses an action sequence by rolling out the model and minimizing terminal error:

Here we use as shorthand for the full rollout through the world model (dependence on model parameters is implicit):

In short horizons and low-dimensional systems, this can work reasonably well. But as horizons grow and models become larger and more expressive, its weaknesses become amplified.

So why doesn’t this just work at scale?

Why long-horizon planning is hard (even when everything is differentiable)

There are two separate pain points for the more general world model, plus a third that is specific to learned, deep learning-based models.

1) Long-horizon rollouts create deep, ill-conditioned computation graphs

Those familiar with backprop through time (BPTT) may notice that we’re differentiating through a model applied to itself repeatedly, which will lead to the exploding/vanishing gradients problem. Namely, if we take derivatives (note we’re differentiating vector-valued functions, resulting in Jacobians that we denote with ) with respect to earlier actions (e.g. ):

We see that the Jacobian’s conditioning scales exponentially with time :

leading to exploding or vanishing gradients.

2) The landscape is non-greedy and full of traps

At short horizons, the greedy solution, where we move straight toward the goal at every step, is often good enough. If you only need to plan a few steps ahead, the optimal trajectory usually doesn’t deviate much from “head toward ” at each step.

As horizons grow, two things happen. First, longer tasks are more likely to require non-greedy behavior: going around a wall, repositioning before pushing, backing up to take a better path. And as horizons grow, more of these non-greedy steps are typically needed. Second, the optimization space itself scales with horizon: , further expanding the space of local minima for the optimization problem.

A long-horizon fix: lifting the dynamics constraint

Suppose we treat the dynamics constraint as a soft constraint, and we instead optimize the following penalty function over both actions and states :

This is also sometimes called collocation in planning/robotics literature. Note the lifted formulation shares the same global minimizers as the original rollout objective (both are zero exactly when the trajectory is dynamically feasible). But the optimization landscapes are very different, and we get two immediate benefits:

• Each world model evaluation depends only on local variables, so all terms can be computed in parallel across time, resulting in a huge speed-up for longer horizons, and
• You no longer backpropagate through a single deep -step composition to get a learning signal, since the previous product of Jacobians now splits into a sum, e.g.:

Being able to optimize states directly also helps with exploration, as we can temporarily navigate through unphysical domains to find the optimal plan:

However, lunch is never free. And indeed, especially for deep learning-based world models, there is a critical issue that makes the above optimization quite difficult in practice.

An issue for deep learning-based world models: sensitivity of state-input gradients

The tl;dr of this section is: directly optimizing states through a deep learning-based is incredibly brittle, à la adversarial robustness. Even if you train your world model in a lower-dimensional state space, the training process for the world model makes unseen state landscapes very sharp, whether it be an unseen state itself or simply a normal/orthogonal direction to the data manifold.

Adversarial robustness and the “dimpled manifold” model

Adversarial robustness originally looked at classification models , and showed that by following the gradient of a particular logit from a base image (not of class ), you did not have to move far along to make classify as (Szegedy et al., 2014; Goodfellow et al., 2015):

Later work has painted a geometric picture for what’s going on: for data near a low-dimensional manifold , the training process controls behavior in tangential directions, but does not regularize behavior in orthogonal directions, thus leading to sensitive behavior (Stutz et al., 2019). Another way stated: has a reasonable Lipschitz constant when considering only tangential directions to the data manifold , but can have very high Lipschitz constants in normal directions. In fact, it often benefits the model to be sharper in these normal directions, so it can fit more complicated functions more precisely.

As a result, such adversarial examples are incredibly common even for a single given model. Further, this is not just a computer vision phenomenon; adversarial examples also appear in LLMs (Wallace et al., 2019) and in RL (Gleave et al., 2019).

While there are methods to train for more adversarially robust models, there is a known trade-off between model performance and adversarial robustness (Tsipras et al., 2019): especially in the presence of many weakly-correlated variables, the model must be sharper to achieve higher performance. Indeed, most modern training algorithms, whether in computer vision or LLMs, do not train adversarial robustness out. Thus, at least until deep learning sees a major regime change, this is a problem we’re stuck with.

Why is adversarial robustness an issue for world model planning?

Consider a single component of the dynamics loss we’re optimizing in the lifted state approach:

Let’s further focus on just the base state:

Since world models are typically trained on state/action trajectories , the state-data manifold for has dimensionality bounded by the action space:

where is some optional space of augmentations (e.g. translations/rotations). Thus, we can typically expect to be much lower than , and thus: it is very easy to find adversarial examples that hack any state to any other desired state.

As a result, the dynamics optimization

feels incredibly “sticky,” as the base points can easily trick into thinking it’s already made its local goal.¹

Our fix

This is where our new planner GRASP comes in. The main observation: while is untrustworthy and adversarial, the action space is usually low-dimensional and exhaustively trained, so is actually reasonable to optimize through and doesn’t suffer from the adversarial robustness issue!

At its core, GRASP builds a first-order lifted state / collocation-based planner that is only dependent on action Jacobians through the world model. We thus exploit the differentiability of learned world models , while not falling victim to the inherent sensitivity of the state Jacobians .

GRASP: Gradient RelAxed Stochastic Planner

As noted before, we start with the collocation planning objective, where we lift the states and relax dynamics into a penalty:

We then make two key additions.

Ingredient 1: Exploration by noising the state iterates

Even with a smoother objective, planning is nonconvex. We introduce exploration by injecting Gaussian noise into the virtual state updates during optimization.

A simple version:

Actions are still updated by non-stochastic descent:

The state noise helps you “hop” between basins in the lifted space, while the actions remain guided by gradients. We found that specifically noising states here (as opposed to actions) finds a good balance of exploration and the ability to find sharper minima.²

Ingredient 2: Reshape gradients: stop brittle state-input gradients, keep action gradients

As discussed, the fragile pathway is the gradient that flows into the state input of the world model, . The most straightforward way to do this initially is to just stop state gradients into directly:

• Let be the same value as , but with gradients stopped.

Define the stop-gradient dynamics loss:

This alone does not work. Notice now states only follow the previous state’s step, without anything forcing the base states to chase the next ones. As a result, there are trivial minima for just stopping at the origin, then only for the final action trying to get to the goal in one step.

Dense goal shaping

We can view the above issue as the goal’s signal being cut off entirely from previous states. One way to fix this is to simply add a dense goal term throughout prediction:

In normal settings this would over-bias towards the greedy solution of straight chasing the goal, but this is balanced in our setting by the stop-gradient dynamics loss’s bias towards feasible dynamics. The final objective is then as follows:

The result is a planning optimization objective that does not have dependence on state gradients.

Periodic “sync”: briefly return to true rollout gradients

The lifted stop-gradient objective is great for fast, guided exploration, but it’s still an approximation of the original serial rollout objective.

So every iterations, GRASP does a short refinement phase:

1. Roll out from using current actions , and take a few small gradient steps on the original serial loss:

The lifted-state optimization still provides the core of the optimization, while this refinement step adds some assistance to keep states and actions grounded towards real trajectories. This refinement step can of course be replaced with a serial planner of your choice (e.g. CEM); the core idea is to still get some of the benefit of the full-path synchronization of serial planners, while still mostly using the benefits of the lifted-state planning.

How GRASP addresses long-range planning

Collocation-based planners offer a natural fix for long-horizon planning, but this optimization is quite difficult through modern world models due to adversarial robustness issues. GRASP proposes a simple solution for a smoother collocation-based planner, alongside stable stochasticity for exploration. As a result, longer-horizon planning ends up not only succeeding more, but also finding such successes faster:

Push-T results. Success rate (%) / median time to success. Bold = best in row. Note the median success time will bias higher with higher success rate; GRASP manages to be faster despite higher success rate.

What’s next?

There is still plenty of work to be done for modern world model planners. We want to exploit the gradient structure of learned world models, and collocation (lifted-state optimization) is a natural approach for long-horizon planning, but it’s crucial to understand typical gradient structure here: smooth and informative action gradients and brittle state gradients. We view GRASP as an initial iteration for such planners.

Extension to diffusion-based world models (deeper latent timesteps can be viewed as smoothed versions of the world model itself), more sophisticated optimizers and noising strategies, and integrating GRASP into either a closed-loop system or RL policy learning for adaptive long-horizon planning are all natural and interesting next steps.

I do genuinely think it’s an exciting time to be working on world model planners. It’s a funny sweet spot where the background literature (planning and control overall) is incredibly mature and well-developed, but the current setting (pure planning optimization over modern, large-scale world models) is still heavily underexplored. But, once we figure out all the right ideas, world model planners will likely become as commonplace as RL.

For more details, read the full paper or visit the project website.

Citation

@article{psenka2026grasp,
  title={Parallel Stochastic Gradient-Based Planning for World Models},
  author={Michael Psenka and Michael Rabbat and Aditi Krishnapriyan and Yann LeCun and Amir Bar},
  year={2026},
  eprint={2602.00475},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2602.00475}
}

This article was initially published on the BAIR blog, and appears here with the authors’ permission.

AIhub is excited to launch a new series, speaking with leading researchers to explore the breakthroughs driving AI and the reality of the future promises – to give you an inside perspective on the headlines. The first interviewee is Ross King, who created the first robot scientist back in 2009. He spoke to us about the nature of scientific discovery, the role AI has to play, and his recent work in DNA computing.

Automated science is a really exciting area, and it feels like everyone’s talking about it at the moment – e.g. AlphaFold sharing the 2024 Nobel Prize. But you’ve been working in this field for many years now. In 2009 you developed Adam, the first robot scientist to generate novel scientific knowledge. Could you tell me some more about that?

So the history goes back to before Adam. Back in the late 1990s, I moved from a postdoc at what was then the Imperial Cancer Research Fund – now Cancer Research UK – and got my first academic job at the University of Wales, Aberystwyth. That’s where I had the original idea of trying to automate scientific research.

Our first publication on this was in 2004. It was a paper about robot scientists, published in Nature. That was the start. We showed that the different steps in the scientific method – forming hypotheses, determining experiments to test them, analysis of the results – could all be individually automated. But the whole cycle wasn’t fully automated, and the AI system didn’t do any novel science at that point.

In 2009, we built the Adam system. Adam was a (physically) large laboratory automation system, combined with AI that could perform full cycles of scientific research, and had knowledge about yeast functional genomics. Adam hypothesised and experimentally confirmed novel scientific knowledge about yeast metabolism, which we manually verified in the lab. 

How has the field evolved since then?

For many years, not much happened. Funding was difficult due to the financial crisis, which made the British Research Councils much more conservative. Before that period, panels would choose the most exciting science. Afterwards, they focused more on what would help Britain financially in the near term.

We couldn’t get funding for many years, and few others were interested. There was some work in symbolic regression – finding interpretable mathematical models to fit phenomena – but not much automation of science. What changed was the general rise of AI. As AI became more prominent, interest picked up, especially after 2017.

What are the potential upsides and downsides of AI scientists? 

I’ll start with the big picture: I think that science is positive for humanity. I think our lives in the 21st century are better than those of kings and queens in the 17th century, when modern science started. We have better food from around the world, beautiful fruits for breakfast, and much better healthcare – a 17th-century dentist was not pleasant. My mobile phone can communicate with billions of people at the touch of a button, and I can fly around the world. These are unbelievably good standards of living for billions of people, not just elites. The application of science to technology has provided this.  Of course there are downsides – pollution, environmental damage – but generally, for humans, I think life is better than in the 17th century. 

However, we still have huge problems. We can’t stop global warming or many diseases, and a billion people still live with food insecurity. I think we have sufficient technology to solve these problems if the nations of the world collaborated and shared resources. But I see no prospect of that happening in the current world situation, and I see no examples from history where these things have happened. So my only hope is that science becomes more efficient. If AI can help achieve that, then perhaps we can overcome these challenges. If we have better technology and we treat people badly after that, then it’s not down to constraints in the world, it’s down to human beings. 

As for having AI scientists as colleagues: AI systems don’t understand the big picture. They can’t do really clever things, like Einstein seeing space and time as a four-dimensional continuum as opposed to quite separate things. If you read the 1905 paper by Einstein, it starts off with this philosophical problem about electricity and magnets – AI systems are nowhere near as clever as being able to do anything like that. They can’t see deep analogies or connections, but they are brilliant at other parts of science. They can literally read everything – they have read every paper in the world 1000 times. If you have a small amount of data, machine learning systems can analyze it better than humans would. In this sense, they have superhuman powers. 

One interesting thing now is that if you’re a working scientist and you’re not using AI, in almost all fields you’re not going to be competitive anymore. AI on its own is not better than humans – yet. But a human plus AI is better than a human alone. Human scientists need to embrace AI and use it to do better science.

Do you think we’ll reach a point where autonomous AI will be able to generate the research questions and direct the movement of research?

Yes, I think so, although we’re not close to that at the moment. They can generate new ideas in constrained spaces, often better than humans, but they don’t really have the big picture yet. 

I think that will come sooner or later. I’m involved in a project called the Nobel Turing Challenge. The goal of that is to build an AI robotic system able to do autonomous science at the level of a Nobel Prize winner, by the year 2050. And if you can do that, we can build two machines, a hundred machines, a million machines – and we’d transform society.

Do you think that’s feasible by 2050? 

Just before the pandemic and during the pandemic, I thought the probability of hitting that target was dropping. But then there was the breakthrough of large language models, which are amazing in many ways – often remarkably stupid too, but generally very clever. I think that they alone will not be enough to beat the Nobel Turing Challenge, but I think they’ve made the probability of hitting that target much more likely.

What is interesting – and I don’t know the answer to this – is whether you need to solve AI in general to solve science, or whether it’s more like chess, where you can build a special machine which is genius at chess but not anything else. Imagine some machine which is a genius at physics but doesn’t know anything about poetry or history. Would that be enough? 

My instinct would be to say that it’s not, because everything’s so interlinked – poetry has rhythm, music contains mathematical structures. I think an AI scientist would need a broader understanding of reality than just its specific domain. 

People used to think that we needed those things to solve chess, so our human intuition is not very good at these things. For example, I didn’t expect LLMs to work so well, just by building a bigger network and putting in more data. I assumed they’d need some deep internal model of the world, or even that they would need a body to really understand how things move around in the world.

LLMs raise some interesting questions – are they just mimicking intelligence, as they lack internal models? 

I think AI must have, in some sense, some internal model inside. It’s just we don’t really understand why they work. It’s purely empirical, which is very unusual. I don’t remember a case where we have such an important technology, but we have so little understanding of it.

It is quite mysterious. Especially because science is always asking “what’s the mechanism?”  With AI, it’s the opposite. The question is “does it work?” We don’t know what the mechanism is. 

It’s not even clear what the theory to explain it is. Coming from machine learning, I assumed it would be some sort of Bayesian inference or something. But the mathematicians say no, it’s all to do with function mapping in some high dimensional space. These don’t seem to be the same, so it’s not even clear what framework we should use to explain it. 

And, mapping in a high dimensional space is something that’s fundamentally not intuitively understandable to humans. 

Yes, so it’s a mystery. So why do they do so well, and why do they not overfit over so many parameters. How do they manage to come to a reasonable answer? Generally, it’s easy to understand why they make mistakes, but it’s not so easy to understand why they actually work so well. 

Can you speak about your work in DNA computing, and how it relates to automated science?

With automated science, we’re using computer science to understand, for instance, biology or chemistry. With DNA computing we’re using technology from biology and chemistry to improve computer science. With DNA, you have the potential to have many, many orders of magnitude greater computing density than with electronics. This is because the bases in DNA are roughly the same size as the smallest transistors, but you can pack DNA in three dimensions, whereas transistors can only be in two dimensions. In our design for DNA, every DNA strand is a tiny computer. 

And the beautiful thing with DNA is that it can replicate itself – nature has made ways of copying DNA which are very effective. That’s how we as humans and all animals and plants and bacteria replicate, whereas electronic computers don’t replicate themselves – they’re built in factories costing billions. We can piggyback on top of this wonderful technology which nature has given us.

How does a DNA computer work? 

One of the greatest discoveries ever made was by Alan Turing, who discovered, or invented, the concept of the universal Turing machine. So this is an abstract mathematical object which can essentially compute anything which any other computer can compute. You can’t make a more powerful computer, in the sense that it can compute a function which that universal Turing machine can’t compute.

And there’s many different ways of physically implementing a universal Turing machine. The most common one is to build an electronic computer. But you could, in principle, build a Turing machine out of tin cans, for instance – the only difference is how fast they go and how much memory they have. The reason that your computer can do multiple tasks is because it can be programmed to do.

The beautiful thing which you can do with DNA is you can make a non deterministic universal Turing machine. These compute the same functions as normal universal Turing machines, but they do so exponentially faster – every time there is a decision point in the program, rather than having to explore only one path, it can go both ways simultaneously. So you can make a computer which, like an organism (think rabbits), can replicate and replicate and replicate until we solve the problem, or you run out of space. So space becomes the limiting factor rather than time. 

You can imagine that if you wanted to search through a tree to find something, you could put down all the branches in parallel, whereas a normal computer would go down one branch at a time. If you do the sums for DNA computing, you could have more memory and more compute on a desktop than all the electronic computers on the planet, which seems incredible. That’s just because of the density of compute. 

That would be an incredible scale-up – like how a modern smartphone is so  much more powerful than NASA’s supercomputers in the 60s. But computing isn’t improving at the same rate as it used to. 

Yes. Computers are not improving like they used to for many decades (Moore’s law). That’s why these big tech companies are building big compute farms the size of Manhattan or soon maybe Texas. So the world does need more efficient ways of doing compute.

If we had a lot of compute, what kinds of scientific problems or areas do you think AI-enabled science could best be applied to? Are there any low-hanging fruits?

What’s very important is to integrate AI systems with actual experiments and laboratories. You can’t just think about science and get the right answer. We need to actually go into the labs and test things, but a lot of AI people and AI companies don’t really appreciate that. They’ve been so successful in science with AI plus simulation that they don’t realize simulation is only so good as something that’s testable.

Areas with low-hanging fruit include materials science, as we need better battery materials, better solar panels, and lots more. There’s something of a gold rush happening there right now, with many startup companies getting huge valuations.

The other area of automation, which is in some sense easier, is drug design, because it’s much easier to move liquids around than solid phase materials. Closed-loop automation has sort of transformed early-stage drug design, and there are lots of companies in that space now.

The big picture is that the economic cost of science is dropping. A lot of the actual thinking involved in science can now be done by AI systems, and the experimental work can be done very well by lab automation. You don’t need to employ people to move things around, and people aren’t as accurate and don’t record things as well as automation does. So that’s the big picture: what can we do if we can make science much cheaper?

Where do you think AI science is headed next?

I think there’s an analogy with computer games like chess and Go. In my lifetime, computers went from playing chess pretty poorly to being able to beat the world champion. I think it’s the same in science. There’s a continuum of ability from what current technology can do, from the average human, to grandmasters of science like Newton, Einstein, Darwin and others. If you agree there is no sharp cutoff on that path, then I think that with faster computers, better algorithms, and better data, there’s nothing stopping them getting better and better at science. Whereas there’s evidence that humans are getting worse at science – the average economic benefit per scientist is decreasing. I think they’ll get better and better and sooner or later overtake humans in science. We shall see, but I’m optimistic. If we get through this period, better science can improve the standard of living and happiness of humanity,  and save the planet at the same time.

And now we have so much data, we need that raw power and intelligence to look at it all.

Yes, we need factories doing a lot of automation to scale things up. There’s no point in AI having brilliant ideas if we can’t test them in the lab. In my mind, science is still at the pre-industrial level. A PI with some post-docs and a few students is like a cottage industry, as opposed to a factory of science. I think humans will still be doing science, but we won’t be actually pipetting things in the future. It’s one reason we chose the name Adam (Adam Smith), we want to change the economics of science. 

And Eve?

Eve was a system we developed some years ago to look at early-stage drug design. Eve optimises a process, rather than doing pure science. Most systems don’t actually do hypothesis-driven science, they optimise something, e.g. find a better material for batteries, which is useful, but not necessarily science. 

Our new system is called Genesis. There we’re trying to scale up the experiments we can do and build up a lot of data. We’re using a continuous flow bioreactor, which enables you to control the growth rate of microorganisms. This is important if you want to understand their internal workings.

And you’re beginning with microorganisms because they’re a fundamental unit of life? 

Yes, we want to understand the eukaryotic cells. There are three branches of life, and the other two are bacteria. Eukaryotes evolved more than 1 billion years ago. We are eukaryotes. Biology is conservative, so the design of yeast and human cells is pretty much the same, but yeast cells are much simpler than human ones. To understand how we work, first we need to understand yeast, then human cells. Once we understand how human cells work, we can understand how organs work, then how humans work, and then we can solve medicine. It’s a reductionist approach to science – we understand something simple first, and then build from there. 

I like the progression, that approach makes sense. 

Unfortunately, it doesn’t make sense to our funders. They generally want to fund practical work on human cells now. They don’t easily fund research on fundamental questions. 

That’s the problem with the funding system. Most great discoveries in science over the last few centuries would not have been funded – they happened because people were doing the most impractical things for the most impractical reasons. And maybe a century later they were found to have a practical purpose. 

Exactly. Some years ago in the UK you had to write a 2-pages for every Research Council grant on how your research was going to make Britain richer or healthier. What would Alan Turing have written on his grant application for the Entscheidungsproblem? 

Thank you. This has been a very interesting conversation.

Thank you, happy to discuss this. It’s a very interesting topic. 

About Ross King

Ross King is a Professor with joint positions at the University of Cambridge, and Chalmers Institute of Technology, Sweden. He originated the idea of a ‘Robot Scientist’: integrating AI and laboratory robotics to physically implement scientific discovery. His research has been published in top scientific journals – Science, Nature, etc. – and received wide publicity. His other core research interest is DNA computing. He developed the first nondeterministic universal Turing machine, and is now working on a DNA computer that can solve larger NP complete problems than conventional or quantum computers.

In the latest in our series of interviews meeting the AAAI/SIGAI Doctoral Consortium participants, we caught up with Aniket Roy to find out more about his research on generative models for computer vision tasks.

Tell us a bit about your PhD – where did you study, and what was the topic of your research?

I recently completed my PhD in Computer Science at Johns Hopkins University, where I worked under the supervision of Bloomberg Distinguished Professor Rama Chellappa. My research primarily focused on developing methods for resource-constrained image generation and visual understanding. In particular, I explored how modern generative models can be adapted to operate efficiently while maintaining strong performance.

During my PhD, I worked broadly at the intersection of generative AI, multimodal learning, and few-shot learning. Much of my work involved designing techniques that enable models to learn new concepts or perform complex visual tasks with limited data or computational resources. This included research on diffusion models, personalized image generation, and multimodal representation learning. Overall, my work aims to make advanced vision and generative AI systems more adaptable, efficient, and practical for real-world applications.

Could you give us an overview of the research you carried out during your PhD?

During my PhD, my research broadly focused on improving the adaptability, efficiency, and quality of modern generative models for computer vision tasks. The rapid progress in generative AI–particularly diffusion models and vision–language models–has created new opportunities to address long-standing challenges such as data scarcity, controllable generation, and personalized image synthesis. My work aimed to develop methods that allow these large models to adapt effectively with limited data and computational resources while maintaining high visual fidelity.

One line of my research addressed learning in data-constrained settings. For example, I proposed FeLMi, a few-shot learning framework that leverages uncertainty-guided hard mixup strategies to improve robustness and generalization when only a small number of labeled samples are available. Building on this idea of improving training data quality, I also developed Cap2Aug, which introduces caption-guided multimodal augmentation. This approach uses textual descriptions to guide synthetic image generation, improving visual diversity while reducing the domain gap between real and generated data.

Overview of Cap2Aug.

Another aspect of my research focused on improving the perceptual quality of images generated by diffusion models. In this direction, I proposed DiffNat, a plug-and-play regularization method based on the kurtosis-concentration property observed in natural images. By incorporating this principle into diffusion models through a KC loss, the generated images exhibit more natural texture statistics and improved perceptual realism, which also benefits downstream vision tasks.

A major part of my work explored personalization and efficient adaptation of large generative models. I introduced DuoLoRA, a parameter-efficient framework for composing low-rank adapters that enables fine-grained control over content and style without requiring full retraining of the base model. I further extended personalization to zero-shot settings using a training-free textual inversion approach that allows arbitrary objects to be customized directly during generation. Finally, I proposed MultiLFG, a frequency-guided multi-LoRA composition framework that uses wavelet-domain representations and timestep-aware weighting to enable accurate and training-free fusion of multiple concepts in diffusion models.

Overview of DuoLoRA.

Overall, my research contributes toward building generative systems that are more efficient, adaptable, and controllable, enabling high-quality image generation and understanding even in data-limited or resource-constrained scenarios.

Was there a specific project or an aspect of your research that was particularly interesting?

One project that I found particularly interesting during my PhD is DiffNat, which was published in TMLR 2025. Diffusion models have become the backbone of many modern generative AI systems and have achieved impressive results in generating and editing realistic images. However, improving the perceptual quality and naturalness of generated images remains an important challenge.

Overview of DiffNat.

In this work, we introduced a simple but effective regularization technique called the kurtosis concentration (KC) loss, which can be integrated into standard diffusion model pipelines as a plug-and-play component. The idea was inspired by a statistical property of natural images: when an image is decomposed into different band-pass filtered versions–for example using the Discrete Wavelet Transform–the kurtosis values across these frequency bands tend to be relatively consistent. In contrast, generated images often show large discrepancies across these bands. Our method reduces the gap between the highest and lowest kurtosis values across the frequency components, encouraging the generated images to follow more natural image statistics.

In addition, we introduced a condition-agnostic perceptual guidance strategy during inference that further improves image fidelity without requiring additional training signals. We evaluated the approach across several diverse tasks, including personalized few-shot finetuning with text guidance, unconditional image generation, image super-resolution, and blind face restoration. Across these tasks, incorporating the KC loss and perceptual guidance consistently improved perceptual quality, measured through metrics such as FID and MUSIQ, as well as through human evaluation.

What I particularly liked about this project is that it connects classical image statistics with modern diffusion models. It shows that relatively simple statistical insights about natural images can still play a powerful role in improving large generative models.

What are your plans for building on the PhD – where are you working now and what will you be investigating next?

During my PhD, I discovered that I genuinely enjoy the process of research–especially the moment when an intuition or idea turns out to work in practice. That process of exploring new ideas and pushing the boundaries of what we know is something I find very motivating.

To continue pursuing this, I will be joining NEC Laboratories America as a Research Scientist. In this role, I hope to build on my PhD work by developing new methods for generative models and exploring how these models can interact with broader multimodal systems. In particular, I am interested in advancing research at the intersection of generative models, vision–language–action models, and embodied AI. More broadly, my goal is to contribute to the development of intelligent systems that can understand, generate, and interact with the visual world more effectively, while also continuing to push forward the scientific understanding of these models.

I’m interested in how you got into the field. What inspired you to study computer vision and machine learning?

My interest in computer vision and machine learning started during my undergraduate studies, when I took courses in signal processing and image processing. I found those subjects particularly fascinating because they allowed you to experiment with algorithms and immediately see their effects on images. That visual and intuitive aspect made the field very engaging, and it helped me appreciate how mathematical concepts can directly translate into meaningful visual results.

At the same time, I was also curious about how the human brain processes visual information—how we are able to recognize objects, understand scenes, and interpret complex visual signals so effortlessly. That curiosity led me to wonder whether we could design computational models that mimic aspects of human perception and enable machines to understand visual data in a similar way.

A major influence during this time was my professor, Dr. Kuntal Ghosh, who encouraged me to think more deeply about these problems and approach them with a scientific mindset. His mentorship played an important role in shaping my interest in research. Since then, that curiosity about visual perception and intelligent systems has continued to drive my work in computer vision and machine learning.

What was your experience of the Doctoral Consortium at AAAI?

Unfortunately, I was not able to attend the AAAI Doctoral Consortium in person due to visa-related issues. However, a colleague kindly helped present my poster on my behalf during the event. Even though I could not be there physically, I was very encouraged by the response my work received. Several researchers reached out to me after seeing the poster, and we had some very insightful discussions about the ideas and potential future directions of the research. In that sense, I still found the experience quite rewarding. The Doctoral Consortium is a great platform for sharing early-stage ideas, receiving feedback from the community, and connecting with other researchers working on related problems. I appreciated the opportunity to engage with people who were interested in the work, and those interactions helped spark new perspectives and collaborations.

Could you tell us an interesting (non-AI related) fact about you?

Outside of research, I’m a big fan of music and stand-up comedy, and I really enjoy traveling whenever I get the chance. Exploring new places, cultures, and perspectives is something I find refreshing—it’s a great way to recharge and stay curious about the world beyond work. I also enjoy writing poetic satire from time to time, and I occasionally perform it. It’s a fun creative outlet that allows me to mix humor and storytelling, which is quite different from the analytical nature of the research work I usually do.

About Aniket Roy

Aniket is currently a Research Scientist at NEC Labs America. He obtained his PhD from the Computer Science dept at Johns Hopkins University under the guidance of Bloomberg Distinguished Professor Prof. Rama Chellappa. Prior to that, he did a Master’s from Indian Institute of Technology Kharagpur. He was recognized with the Best Paper Award at IWDW 2016 and the Markose Thomas Memorial Award for the best research paper at the Master’s level. During PhD, he explored domains of few-shot learning, multimodal learning, diffusion models, LLMs, LoRA merging with publications in leading venues such as NeurIPS, ICCV, TMLR, WACV, CVPR and also 3 US patents filed. During his PhD, he also gained industrial experience through multiple internships in Amazon, Qualcomm, MERL, and SRI International. He was awarded as an Amazon Fellow (2023-24) at JHU and selected to participate in ICCV’25 and AAAI’26 doctoral consortium.

By Anthony King

What if training a robot to handle dirty, dangerous work on the factory floor was as simple as showing it how? Czech startup RoboTwin is doing exactly that, helping factory workers teach robots new skills by demonstration.

Instead of writing complex code, workers perform the job once and RoboTwin’s technology turns those movements into a robot programme – opening the door to automation for smaller manufacturers.

Founded in Prague in 2021, RoboTwin builds handheld devices and no-code software that capture human movements and translate them into instructions for industrial robots. The aim is to make automation faster, simpler and more accessible to manufacturers that do not have specialist robotics programmers.

“The robot basically copies the human demonstration,” said Megi Mejdrechová, RoboTwin’s co-founder and chief technology officer. “People with no coding skills can transfer their know-how and experience to robots.”

Mejdrechová, a mechanical engineer trained at the Czech Technical University in Prague, developed the core technology behind RoboTwin during her work in robotics research and industry. Her experience in robot control using AI and computer vision inspired her to create something practical for European manufacturers.

“Czech engineering is quite traditional and focused on scientific papers,” said Mejdrechová. “Visits to Singapore and Canada and other work experiences led me to focus on making a product that people could use.”

Getting started

In 2021, Mejdrechová entered a jump‑starter programme and won first prize in the manufacturing category. “We saw then that there was potential for the technology,” she said.

This encouraged her to start RoboTwin with colleagues Ladislav Dvořák and David Polák, who shared her enthusiasm for human‑robot partnerships. Mejdrechová received backing from Women TechEU, an EU scheme supporting women founders of deep‑tech startups. 

The RoboTwin team shared their results on the Horizon Results Platform, an online showcase for EU‑funded innovations, which led to an invitation to the EU’s Empowering Start‑ups and SMEs initiative. 

This helped fund their trip to Hannover Messe 2025, a major global manufacturing trade fair, and opened doors to new business contacts and deals.

Through a mix of public and private investment, RoboTwin has secured funding to refine its technology and expand to manufacturers in Central Europe, the Netherlands, Mexico and Canada.

In 2025, Mejdrechová was named in Forbes Czechia’s 30 Under 30 list for her work in making the training of robots accessible to more manufacturers.

Schooling robots

At the heart of RoboTwin’s system is a handheld device equipped with sensors. When a worker performs a task, for example spray painting a metal component, the system records the movement and converts it into a robot programme that can be reused in production.

Instead of requiring a specialist engineer to manually code every movement, the system captures the worker’s natural technique and translates it into precise instructions a robot can follow.

“We started with jobs that are ugly, dirty and unhealthy for workers to do manually,” said Mejdrechová.

Thanks to the no‑coding system, the process can be completed in a few steps and typically takes about a minute. For factories producing small batches or frequently changing products, this speed can make automation far more practical than traditional robot programming.

Making automation easy for all

Robotics in manufacturing is not new. The automotive industry already leads the way with about 23 000 new robots added to production lines in 2024. But while large companies can invest heavily, automation remains challenging and expensive for many SMEs.

This is where RoboTwin lends a hand. It has assisted firms in the surface‑treatment industry – companies that powder coat, paint or polish metal or plastic parts for car factories.

“Even if the batch of products you are producing is small, with our approach you can create a robot programme fast and easily,” said Mejdrechová.

For example, RoboTwin has assisted RobPainting, a Dutch company that robotises painting for SMEs to improve quality, reduce costs and minimise rework. 

“With our device we can teach the robot precise trajectories that are needed for a product and about its surroundings,” said David Vobr, a robotics specialist at RoboTwin who often assists customers.

Dangerous jobs

RoboTwin’s system can work with a wide range of industrial robots, including collaborative robots designed to operate safely alongside humans.

“We can have manipulators or painting robots and also collaborative robots, which can work alongside humans because they have sensors that tell them when to stop moving if someone could get hurt,” said Vobr.

RoboTwin initially focused on surface treatment in manufacturing, where tasks such as spray painting require workers to wear protective clothing and perform repetitive movements.

“These jobs are difficult to automate because there is often a lot of hidden movement involved,” said Mejdrechová, referring to small adjustments and gestures that workers make instinctively.

The sector also faces labour shortages.

“People are often not happy doing these things and there is a lack of workers willing to take these jobs. So there is a high demand for automation.”

Customers report that many robot programmes can now be created without shutting down a production line.

RoboTwin has already worked with a number of companies, including Surfin Technology, a Czech company specialising in robotic coating solutions, and Innovative Finishing Solutions in Canada, which brings its technology to North American customers.

Scaling up

EU support for RoboTwin is ongoing. A €2.3 million grant from the European Innovation Council secured in 2025 will help accelerate product development and market expansion.

The funding will support the next generation of RoboTwin’s technology. Instead of relying solely on manual demonstrations, the system will increasingly use stored experience and data to generate robot programmes automatically based on the shape of an object.

The company says this could make automation viable for many manufacturing tasks that were previously too complex or costly to automate.

For Europe, technologies like RoboTwin could play an important role in strengthening digital sovereignty and smart industrial innovation. They can help smaller manufacturers adopt advanced robotics without needing specialised programming expertise.

As factories become more flexible and data-driven, the ability to quickly teach robots new tasks may prove increasingly valuable.

Mejdrechová believes this shift will help bring automation within reach of a much wider range of companies.

“Our goal is to make robot training something that workers can do themselves,” she said. “If we succeed, automation will no longer be limited to large factories with specialised engineers. It will become a tool that any manufacturer can use.”

Research in this article was funded by the EU’s Horizon Programme. The views of the interviewees don’t necessarily reflect those of the European Commission. If you liked this article, please consider sharing it on social media.

This article was originally published in Horizon, the EU Research and Innovation magazine.

By Celia Luterbacher

If the goal of a robot is to perform a function, then minimizing the possibility of failure is a top priority when it comes to robotic design. But this minimization is at odds with the robotic raison d’être: systems with multiple units, or agents, can perform more diverse functions, but they also have more different parts that can potentially fail.

Researchers led by Jamie Paik, head of the Reconfigurable Robotics Laboratory (RRL) in EPFL’s School of Engineering, have not only circumvented this problem, but flipped it: they have designed a modular robot that actually lowers its odds of failure by sharing resources among its individual agents.

“For the first time, we have found a way to reverse the trend of increasing odds of failure with increasing function,” Paik explains. “We introduce local resource sharing as a new paradigm in robotics, reducing the failure rate with a larger number of modules.”

In a paper published in Science Robotics, the team showed how exploiting redundant resources and sharing them locally enabled a modular origami robot to successfully navigate a complex terrain, even when one module was completely deprived of power, sensing, and wireless communication.

Sharing is caring

The RRL team took inspiration for their innovation from nature, where the problem of failure is often solved collectively. Birds share local sensing information through flocking behavior, some trees communicate threats to neighbors using airborne signals, and cells continuously transport nutrients across their membranes so that the death of any individual doesn’t significantly impact the overall organism.

Modular robots, which are composed of multiple units that connect to form a complete system, are analogous to multicellular or collective organisms, but until now, their design has been a source of vulnerability: the failure of one module often disables some, if not all, of the robot’s ability to perform tasks. Some modular robots get around this problem with built-in backup resources or self-reconfiguration abilities, but these approaches usually don’t completely restore functionality.

For their study, the RRL team used something called hyper-redundancy: the sharing of all critical power, communication, and sensing resources across all modules, without any change to the robot’s physical structure.

“We found that sharing just one or two resources was not enough: if each resource had an equal chance of failure, system reliability would continue to drop with an increasing number of agents. But when all resources were shared, this this trend was reversed,” Paik says.

In a locomotion task experiment with the Mori3 robot, which is composed of four triangular modules, the team experimented with cutting battery power, wireless communication, and sensing to the central module. Normally, this ‘dead’ central module would block the articulation and movement of the other three, but thanks to hyper-redundancy, the neighboring modules fully compensated for its lack of resources. This allowed the Mori3 to successfully ‘walk’ toward a barrier and contort itself effectively to pass underneath it.

“Essentially, our methodology allowed us to ‘revive’ a dead module in a collective and bring it back to full functionality. Our local resource-sharing framework therefore has the potential to support highly adaptive robots that can operate with unprecedented reliability, finally resolving the reliability-adaptability conflict,” summarizes RRL researcher and first author Kevin Holdcroft.

The researchers say that future work could focus on applying their resource sharing framework to more complex systems with increasing numbers of agents. In particular, the same concept could be extended to robotic swarms, with hardware adaptations that allow swarm members to dock to each other for energy and information transfer.

References

Scalable robot collective resilience by sharing resources, Holdcroft, K., Bolotnikova, A., Monforte, A.J., and Paik, J., Science Robotics (2026).

RoboCup is an international competition that promotes and advances robotics and AI through the challenges presented by its various leagues. We got the chance to sit down with Professor Manuela Veloso, one of RoboCup’s founders, to find out more about how it all started, how the community has grown over the years, and the vision for the future.

I think it would be very interesting to go right back to the beginning and hear how RoboCup got started. What was the initial idea, and how did it get set up?

So we are talking about the mid-90s. In terms of the research in those days, it was the beginning of the internet and many AI and computer science researchers were focused on the internet, first on sophisticated search algorithms, on natural language understanding, on information retrieval, and then on software agents and machine learning applied to digital information. From what I recall, there was a smaller group of researchers who were interested in actual, physical robots, and in particular in AI and robotics. I myself was specifically interested in the problem of creating autonomous robots with perception (get information from the world), cognition (select action to achieve goals), and then act (execute the planned actions). This combination of perception, cognition, and action is a very good framework for autonomous robots, because they have to get their information from their sensors, they have to reason about actions to achieve their goals, and then execute them. So, during the 90s, I was at Carnegie Mellon with this AI research goal of integrating perception, cognition, and action in autonomous robots.

Over in Canada there was Alan Mackworth, who jointly with his wonderful student, Michael Sahota, built a one-on-one little autonomous robot soccer set-up. Two robots ran on a small field that had a camera overhead, and aimed at scoring each in one of the two little goals. This work showed that this task of kicking a ball and defending and aiming at a goal could be done autonomously. So it was a tremendous demonstration that a robot soccer world could exist. Mostly at the same time, in Japan, Minoru Asada was showing that a big robot could learn with reinforcement learning how to push a ball into a goal. So you have these one-on-one, fully autonomous little robot cars that were pushing balls around in Canada, and then there was this effort of learning to score with a larger robot in Japan. The learning robot didn’t have a team, it was not a real game, but it was showing that reinforcement learning could learn the skill of aiming into a goal. And then there was also Hiroaki Kitano at Sony who was very interested in little humanoids.

So this is very beautiful because all these things came into play – all of us had different interests.

Alan Mackworth did not get involved with RoboCup, but he gave a demonstration of these one-on-one robots at AAAI in 1994. And in those days, I had a PhD student who had just joined – Peter Stone. And Peter was a serious and passionate soccer player. He saw this little game and he came to me and said, “this is what I want to do for my thesis research, robot soccer!” And for me, I was trying to find a research environment where autonomy was needed in the robot world. I had already a student, Karen Haigh, who was working with autonomous office robots, and learning to plan and execute. But with these soccer robots and Peter Stone’s interest everything came to play, and we started robot soccer research in my lab.

In 1996, there was also a robot soccer effort in South Korea, called MIROSOT, and that’s the first competition we participated in. So Peter Stone, myself, and the team we built at Carnegie Mellon – Sorin Achim and Kwun Han – went to South Korea to participate. From South Korea, we flew to a robot soccer workshop in Japan organized by Minoru Asada and Hiroaki Kitano. Also in attendance were Dominique Duhaut, Itsuki Noda, Silvia Coradeschi, and Enrico Pagello. And that’s where RoboCup really started – we decided to do a competition. And the good thing was that Kitano was the chair of IJCAI, which was going to happen in Osaka in the summer of 1997. So we are there in Osaka and literally we came up with this idea of having a robot soccer competition, RoboCup. It was a big moment for us as researchers. We had to come up with the rules of this competition so that people would be able to participate seven months later. We came up with the three leagues that we were interested in and had expertise in.

The small size league, building upon our Carnegie Mellon interests, would have a field with a camera overhead connected to a computer and then the computer would remotely control the robots through radio.

Then Minoru Asada had these bigger robots with wheels, and we created a league that we call the middle-size league to include the robotics research of Minoru Asada and others. And then Itsuki Noda was interested in creating a simulation environment. We thought that this would help get more people participating in this task of robot soccer.

So that’s how the three leagues started: the small size, the middle size, and the simulation. Hiroaki Kitano, Dominique Duhaut and I were in charge of the small-size league, Minoru Asada was in charge of the middle-size league, and Itsuki Noda ran the simulation league.

One of the challenges was to come up with the rules and define the robots and playing fields. I remember my own pragmatism in suggesting that we play on a ping-pong table for the small-size, as a ping-pong table is something that exists in the whole world. That meant that we would have a playing surface, with defined size and texture, anywhere in the world. We decided one ping-pong table for the small-size league, and nine ping-pong tables for the middle-size league.

In the summer of 1997, at IJCAI, when we all went to the actual first RoboCup competition, the space was gorgeous. Hiroaki Kitano had made these beautiful fields and white bleachers around the fields. It was a very beautiful space with an area with computers for the simulation league. There were 80 teams that had joined the simulation online. We were, I believe, five teams for the small size league and about eight to ten for the middle size league.

That’s how it all started. And 1997 it was in Osaka, in 1998 Dominique Duhaut organized RoboCup in Paris, at a modern Science Center, La Villette. And then in 1999, RoboCup was organized by Silvia Coradeschi, in Stockholm, again co-located with the IJCAI conference. In these three years 1997, ’98, and ’99, there were only these three leagues, small size, middle size, and simulation. It was the foundation of everything. RoboCup grew consistently every year in terms of the number of teams, the number of participants, and the number of participating countries.

Some of the pitches in the main soccer arena at RoboCup 2024, held in Eindhoven, The Netherlands.

So how did RoboCup expand and could you talk about the decisions to add extra leagues?

Well, nothing was necessarily planned, it was more related to the research interests that people had. So when we were in Melbourne in 2000, there were two things that were added. One was RoboCupJunior. There was a professor, Elizabeth Sklar, who had a tremendous interest in robotics education for children. She proposed the RoboCupJunior competitions for children K-12. The goal of RoboCupJunior was to train all these young people to do research in robotics. It was and it still is extremely successful. In Melbourne there were probably the same number of children as there were people competing in what we now call the major leagues. As well as the soccer leagues, Elizabeth also ran a dance competition. This was very impressive – the children would dance on stage with their designed, built, and programmed mobile robots.

I believe that the rescue league was also introduced in 2000. The reason why we came up with rescue was because in those days, there was a lot of research on robots in disaster environments. So many people in our groups had an interest in developing robots that were able to handle disaster environments. So, we included that interest. And later on, @home was also introduced because people had an interest in service robots.

The logistics league came later to cover robots inside factories. The reason for introducing new leagues was always to include the community. And we didn’t want anyone to feel they could not come to RoboCup because they did different research with their robots. And so it was a very intellectually inclusive environment for AI and robotics researchers that could perform tasks autonomously – with perception, cognition, and action mostly in teams for particular tasks.

Rescue league arena at RoboCup 2024.

Could you talk about the changes that were announced last year, and the decision behind those?

Back in 1997, Hiroaki Kitano came up with this goal of a robot soccer team being able to beat the human World Cup winners by 2025. (I later tried, in the early 2000s, to rephrase that goal so that it was not about the robots beating the humans, but instead having robots playing alongside human players by 2050.) However, for this to happen, we have started the humanoid league so that the robots would have legs, not wheels. Our roadmap hence moves towards humanoid robots. It also happens that currently there are several robot companies who are producing humanoid robots that are easy to acquire, so researchers do not have the design and build their own humanoid robots, which was a difficult task that was pursued only by a limited group of researchers. In parallel, events started aiming at including humanoid robot soccer. We believed that RoboCup should be visible and a reference for humanoid robot soccer. We then proposed and changed that, in the international RoboCup competitions, robot soccer will focus solely on humanoid robots. That doesn’t mean that the other RoboCup leagues aren’t valuable. But if they continue at a regional level, there will still be venues to foster these other types of interests. The RoboCup international competitions would be solely focused on humanoid robot soccer.

Action from the humanoid soccer league at RoboCup 2025. Image courtesy of Alessandra Rossi.

RoboCup was tremendously successful in terms of including a variety of different autonomous robot research interests from the community. But I also believe that now we should move into just humanoid soccer at the international level. This will help with visibility of the competition and also for consolidation. We can now buy humanoid robots that we didn’t have before. Previously, you had to build them in your universities, and that was not a research direction for many people. But there is not that excuse anymore. You can buy humanoid robots, and increasingly, you can talk with these humanoid robots using GenAI. So it became a different challenge, a much more accessible challenge than before. The challenge will be: can these platforms play soccer in the presence of another team? I think it’s a fascinating new direction to try to focus on this challenge of the game of soccer – multiple players, two teams in a large space, eventually coordinating, collaborating, or playing with and against humans. So that’s the rationale.

When people don’t know the story, they tend to wonder why we have all these other leagues, such as rescue and logistics. It was always driven by the interests of the community and a big heart for a research approach. And we always thought that because we were organizing an event with a venue, that we could include these people that have interests outside of soccer. That has always been our drive. However, it’s true that it diluted a little bit the goal of 2050 – the humanoid soccer robots. Now I think it’s time to go back to soccer as a main focus, and still keep supporting the other interests that we have fostered, more at a regional level.

There are still a lot of details to be sorted out. And, of course, people may be upset by leagues terminating at the international level. I have always participated in the small-size league myself, 20 years of participating, and I am very attached to it. I was always the trustee representing the small-size league, and I am not sure I am happy with the small-size not being in the international event. But I need to think beyond what my personal interests are and try to understand what would have a bigger impact for AI, robotics, and RoboCup. I think that that’s demonstrating a group of humanoids playing soccer. And if we don’t do it at RoboCup, maybe someone else will.

The way I think about it is that we have made this decision and we can reevaluate it in five or ten years. But not making a change and never having the courage to deviate from what we have been pursuing is less exciting. Things have stalled a bit, with the same number of teams, the same people, the same rules, the same type of intellectual and research accomplishments. We need to excite people about something major again. So I think that the community will greatly embrace this new RoboCup international. And we are so well organized locally, that we can support other leagues at the local events.

I do think that, from a scientific and research point of view, it’s the right moment to target the soccer humanoid robots, because of their availability and our RoboCup ultimate goal. I think it would be amazing if people knew RoboCup international as the humanoid robot soccer competition. Can you imagine in 2027 having 100 teams all playing soccer with humanoids? We’ll see.

About Manuela Veloso

Herbert A. Simon University Professor Emerita, School of Computer Science, Carnegie Mellon University 2018-2026 Founder and Head, JPMorganChase AI Research, Managing Director

For the last eight years, Manuela Veloso has been the founder and Head of JPMorganChase AI Research and Herbert A. Simon University Professor Emerita at Carnegie Mellon University, where she was faculty in the Computer Science Department and then Head of the Machine Learning Department.

At JPMorganChase, she built a team of 100 top talented members with graduate education (PhD and Masters) in AI and related disciplines. The team focused on pillar areas of AI in finance, including data-driven optimization, planning and search, document analysis, trustworthy AI, AI and mathematical reasoning, continual learning, and multiagent systems. The team published their research in academic venues and addressed and contributed to business needs and vision.

Veloso has a licenciatura degree in Electrical Engineering and an M.Sc. in Electrical and Computer Engineering from Instituto Superior Técnico, Lisbon, an M.A. in Computer Science from Boston University, and a Ph.D. in Computer Science from Carnegie Mellon University. Veloso has Doctorate Honoris Causa degrees from the Örebro University, Sweden, the Instituto Universitário de Lisboa (ISCTE), Portugal, the Université de Bordeaux, France, and the Universidade Católica of Portugal.

She served as president of the Association for the Advancement of Artificial Intelligence (AAAI), and she is co-founder and a Past President of the RoboCup Federation. She is a fellow of main professional organizations in her area, namely AAAI, IEEE, AAAS, and ACM. She is the recipient of the ACM/SIGART Autonomous Agents Research Award, the Einstein Chair of the Chinese Academy of Sciences, an NSF Career Award, and the Allen Newell Medal for Excellence in Research. Veloso is a member of the National Academy of Engineering with a citation “for contributions to artificial intelligence and its applications in robotics and the financial services industry.” She is also a member of the Academy of Sciences of Portugal.

Her research interests are in AI, including Multiagent Systems, Autonomous Robots, Continual Learning Agents, and AI in Finance. For further details, see Manuela’s webpage.

Robots are becoming increasingly capable in vision and movement, yet touch remains one of their major weaknesses. Now, researchers have developed a miniature tactile sensor that could give robots something much closer to a human sense of touch.

The technology, developed by researchers at the University of Cambridge, is based on liquid metal composites and graphene – a two-dimensional form of carbon. The ‘skin’ allows robots to detect not just how hard they are pressing on an object, but also the direction of applied forces, whether an object is slipping, and even how rough a surface is, at a scale small enough to rival the spatial resolution of human fingertips. Their results are reported in the journal Nature Materials.

Human fingers rely on multiple types of mechanoreceptors to sense pressure, force, vibration, and texture simultaneously. Reproducing this level of multidimensional tactile perception in artificial systems is a significant challenge, especially in devices that are both small and durable enough for practical use.

“Most existing tactile sensors are either too bulky, too fragile, too complex to manufacture or unable to accurately distinguish between normal and tangential forces,” said Professor Tawfique Hasan from the Cambridge Graphene Centre, who led the research. “This has been a major barrier to achieving truly dexterous robotic manipulation.”

To overcome this, the research team developed a soft, flexible composite material, combining graphene sheets, deformable metal microdroplets, and nickel particles, embedded in a silicone matrix.

Inspired by the microstructures found in human skin, the researchers shaped the material into tiny pyramids, some as small as 200 micrometres across. These pyramid structures concentrate stress at their tips, enabling the sensor to detect extremely small forces while maintaining a wide measurement range.

The result is a tactile sensor sensitive enough to detect a grain of sand. Compared with existing flexible tactile sensors, the new device improves size and detection limits by roughly an order of magnitude.

The sensor can also distinguish shear forces from normal pressure, a capability that allows it to detect when an object begins to slip. By measuring signals from four electrodes beneath each pyramid, the sensor can mathematically reconstruct the full three-dimensional force vector in real time.

In demonstrations, the team integrated the sensors into robotic grippers. The robots were able to grasp fragile objects, such as thin paper tubes, without crushing them. Unlike conventional force sensors, which rely on prior information about an object’s properties, the new system adapts in real time through slip detection.

At even smaller scales, microsensor arrays could identify the mass, geometry, and material density of tiny metal spheres by analysing force magnitude and direction. This opens the door to applications in minimally invasive surgery or microrobotics, where conventional force sensors are far too large.

Beyond robotics, the technology could have significant implications for prosthetics. Advanced artificial limbs increasingly rely on tactile feedback to provide users with a sense of touch. Highly sensitive, miniaturised 3D force sensors could enable more natural interactions with objects, improving control, safety, and user confidence.

“Our approach shows that bulky mechanical structures or complex optics are not required to achieve high-resolution 3D tactile sensing,” said lead author Dr Guolin Yun, a former Royal Society Newton International Fellow at Cambridge, and now Professor at the University of Science and Technology of China. “By combining smart materials with skin-inspired structures, we achieve performance that comes remarkably close to human touch.”

Looking ahead, the researchers believe the sensors could be miniaturised even further, potentially below 50 micrometres, approaching the density of mechanoreceptors in human skin. Future versions may also integrate temperature and humidity sensing, moving closer to a fully multimodal artificial skin.

As robots increasingly move out of controlled factory environments and into homes, hospitals, and unpredictable real-world settings, such advances in touch could be transformative — allowing machines not just to see and act, but to truly feel.

A patent application has been filed through Cambridge Enterprise, the University’s innovation arm. The research was supported by the Royal Society, the Henry Royce Institute, and the Advanced Research and Invention Agency (ARIA). Tawfique Hasan is a Fellow of Churchill College, Cambridge.

Reference

Multiscale-structured miniaturized 3D force sensors, Guolin Yun, Zesheng Chen, Zhuo Chen, Jinrui Chen, Binghan Zhou, Mingfei Xiao, Michael Stevens, Manish Chhowalla & Tawfique Hasan, Nature Materials (2026).

By Anthony King

Modern surgery has gone from long incisions to tiny cuts guided by robots and AI. In the process, however, surgeons have lost something vital: the chance to feel inside the body directly. Without palpation, it becomes harder to detect tissue abnormalities during an operation.

A group of surgeons and engineers across Europe is now trying to bring back this vital aspect of surgery.

Working within an EU-funded research collaboration called PALPABLE, they are developing a soft robotic “fingertip” that can sense how firm or soft tissue is during minimally invasive and robotic surgery. The research runs until the end of 2026, with a first prototype expected to be tested by surgeons around March 2026.

By combining optical sensing, soft robotics and AI, the team is designing a probe that mimics the way a fingertip presses and feels during surgery. It would gently probe organs and create a visual map of tissue stiffness, displayed on a screen to guide surgeons as they operate.

Losing the surgeon’s touch

For many surgeons, the loss of direct touch has been one of the quiet trade-offs of modern surgery.

Sooner or later, I believe the vast majority of surgeries will be robotic.

“We started 30 years ago with open surgery and using our fingers,” said Professor Alberto Arezzo from the University of Turin, Italy. He specialises in minimally invasive and robotic surgery and mostly treats patients with colorectal cancer.

“Then we moved into the era of keyhole surgery, which reduced tactile feedback because we began to use long instruments,” he said.

From the 1990s, keyhole surgery became increasingly common, allowing surgeons to operate through small incisions with the help of a camera. Patients benefited from less trauma, shorter hospital stays and faster recovery.

But this came at the expense of physical touch. That matters because tumours often feel different from healthy tissue – stiffer, less pliable or irregular – important differences that experienced hands can detect.

Finding tumour margins

When operating on cancer, surgeons walk a fine line: remove too much tissue and function may suffer; remove too little and cancer may remain, and then spread again, requiring more surgery. 

“We don’t want to do that. We want it done in one shot,” said Dr Gadi Marom at Hadassah Medical Centre in Jerusalem, one of the clinicians involved in the research, who specialises in minimally invasive and robotic surgery on patients with stomach and oesophagus diseases. 

This is where sensing technology could help. By translating physical contact into visual information, such as a colour-coded map showing softer and firmer areas, surgeons could regain a functional equivalent of touch. 

“With a new instrument, we want to be able to determine the margins around a tumour,” said Marom.

Using light to feel

To do that, engineers on the team are turning to light.

The probe they are developing contains fibre-optic cables embedded in a soft, flexible tip. When pressed against tissue, the tip deforms and the light travelling through the fibres changes.

“A silicone dome presses against soft tissue, allowing us to map both the direction and the magnitude of the applied force,” explained Dr Georgios Violakis at Hellenic Mediterranean University in Heraklion, Crete. 

Those tiny shifts in light intensity and wavelength are then translated into information about tissue stiffness.

In the lab, the team has already built and calibrated early versions of the soft membrane and light-based sensors, with partners contributing across the system.

Queen Mary University of London (UK) is helping design and refine the membranes, the Fraunhofer Institute (Germany) is developing the functional films, while Bendabl (Greece), Tech Hive Labs (Greece) and the University of Essex (UK) are advancing the software needed to visualise stiffness and tactile maps.

The prototype will be validated in lab tests before it is used on patients.

The bottom line is that we will be able to give better care to our patients.

The fibre‑optic cables are each about the width of a human hair. Similar sensing technology has long been used to detect small movements in large structures such as aircraft, skyscrapers and nuclear reactors. Here it is being applied on a much smaller scale to detect subtle differences in human tissue. 

“For touching organs inside an anaesthetised patient, the device needs to be both highly accurate and high resolution,” said Professor Panagiotis Polygerinos, a soft robotics researcher at Hellenic Mediterranean University. 

“Something like this might have been possible sooner, but the technology would have been far more expensive and less precise, making it impractical for clinical use.” 

Bringing touch to robots

As surgery grows increasingly robotic, the loss of tactile feedback is becoming more pressing – and restoring a sense of touch even more vital.

“When I operate with a robot I have the advantage of 3D vision,” said Marom. “And I don’t have to stand for the entire surgery.” That matters in long procedures, such as removing a patient’s oesophagus, which can take up to eight hours.

Robotic surgery also raises new possibilities. Marom hopes it may eventually allow surgeons, in carefully selected cases, to remove small tumours from the oesophagus without removing the entire organ.

But there is a downside.

“In robotic surgery, tactile feedback is largely absent,” said Arezzo. “That’s why this work is so important.”

Both surgeons believe robotics will continue to expand in operating theatres, but only if surgeons are given better sensory information.

“Sooner or later, I believe the vast majority of surgeries will be robotic,” said Arezzo.

For Marom, working closely with engineers has been essential. “I am exposed to soft robotics and many new technologies,” he said. “I see how new instruments can be developed.”

“The bottom line is that we will be able to give better care to our patients,” he added.

Research in this article was funded by the EU’s Horizon Programme. The views of the interviewees don’t necessarily reflect those of the European Commission. If you liked this article, please consider sharing it on social media.

This article was originally published in Horizon, the EU Research and Innovation magazine.

By Eduardo B. Sandoval, UNSW Sydney

Last year, Norwegian-US tech company 1X announced a strange new product: “the world’s first consumer-ready humanoid robot designed to transform life at home”.

Standing 168 centimetres tall and weighing in at 30 kilograms, the US$20,000 Neo bot promises to automate common household chores such as folding laundry and loading the dishwasher.

Neo has a built-in artificial intelligence (AI) system, but for tricky tasks it requires a 1X employee wearing a virtual reality helmet to remotely take over the robot. The operator can see whatever the bot does inside your house, and the process is recorded for future learning.

Other household androids are expected to hit the market this year. But Neo shows the issues at play, which will be familiar to anyone who has watched the AI boom of the past few years: products launched with great fanfare and limited capabilities, concealed privacy risks, and invisible remote workers behind the scenes.

The dream of human-like robots

Machines made in the human likeness have figured in mythology and history for millennia.

But the idea they might realistically be practical consumer products is more recent. Yet it’s a popular one: more than 50 companies around the world are developing this type of robot.

Why now? The past few years have seen improvements in hardware such as batteries, motors and sensors – many thanks to the burgeoning electric vehicle industry. At the same time, the AI systems to control the hardware have also become far more capable.

Hurdles remain

Despite huge technical progress, these robots are still clumsy at handling everyday tasks in homes or hospitals or other uncontrolled environments. While specialised bots such as vacuum cleaners have become a familiar sight, the fact remains that human homes aren’t designed for robots.

And for many fiddly tasks, such as folding laundry, more specialised machines do a better job.

To improve performance, the robots will need a lot of real-world data. The best way to gather that data is by putting these mechanical servants to work in actual homes. And the data in question will include a lot of intimate detail about the lives of specific people – which raises big questions about privacy.

And behind the scenes, at least for now, will be humans. Remote online labour in the tech industry is a growing phenomenon that can increase socioeconomic inequality and have a negative impact on people in developing countries working long hours for low pay, often exposed to disturbing scenes and content.

Other uses for humanoid bots

According to the International Federation of Robotics, useful and widely accepted home androids may still be 20 years away.

But there are other reasons we might want to make artificial humanoids. Japanese researcher Hiroshi Ishiguro has been making human-like “geminoids” for decades with quite different motivations.

My motivation for making humanoid robots stems from an interest in understanding what makes us human, and what it means to be human.

From this perspective, humanoid robots can serve the philosophical exploration of human identity, rather than making life more convenient or generating profits.

What’s ahead

Autonomous humanoid robots will undoubtedly improve as products with the integration of large language models and other generative AI systems.

In the long term, dexterity, navigation, learning and autonomy will get better – but that will require years of research and investment. Humanoid robots will not be immediately available as convincing and useful commercial products.

Concerns around remote work may fade, too. Just last week, 1X announced a software update for its robots that it says will mean less human involvement behind the scenes.

Privacy concerns seem an inherent risk of the technology. An incredibly sophisticated robot in your home will inevitably collect intimate data about your life, opening a new frontier for data exploitation and potential breaches.

Despite these issues, humanoid robots will keep inspiring scientists, engineers and designers. By all means let them inspire us – but we should think twice before letting them stack our dishwashers.

Eduardo B. Sandoval, Scientia Researcher, Social Robotics, UNSW Sydney

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Gerard Dooly, University of Limerick

Plastic pollution is one of those problems everyone can see, yet few know how to tackle it effectively. I grew up walking the beaches around Tramore in County Waterford, Ireland, where plastic debris has always been part of the coastline, including bottles, fragments of fishing gear and food packaging.

According to the UN, every year 19-23 million tonnes of plastic lands up in lakes, rivers and seas, and it has a huge impact on ecosystems, creating pollution and damaging animal habitats.

Community groups do tremendous work cleaning these beaches, but they’re essentially walking blind, guessing where plastic accumulates, missing hot spots, repeating the same stretches while problem areas may go untouched.

Years later, working in marine robotics at the University of Limerick, I began developing tools to support marine clean-up and help communities find plastic pollution along our coastline.

The question seemed straightforward: could we use drones to show people exactly where the plastic is? And could we turn finding the plastic littered on beaches and cleaning it up into something people enjoy – in other words, “gamify” it? Could we also build on other ways that drones have been used previously such as tracking wildfires or identifying shipwrecks.

Building the technology

At the University of Limerick’s Centre for Robotics and Intelligent Systems, my team combined drone-based aerial surveillance work with machine-learning algorithms (a type of artificial intelligence) to map where plastic was being littered, and this paired with a free mobile app that provides volunteers with precise GPS coordinates for targeted clean-up.

The technical challenge was more complex than it appeared. Training computer vision models to detect a bottle cap from 30 metres altitude, while distinguishing it from similar objects like seaweed, driftwood, shells and weathered rocks, required extensive field testing and checks of the accuracy of the detection system.

The development hasn’t been straightforward. Early versions of the algorithm struggled with shadows and confused driftwood for plastic bottles. We spent months refining the system through trial and error on beaches around Clare and Galway so the system can now spot plastic as small as 1cm.

We conducted hundreds of test flights across Irish coastlines under varying environmental conditions, different lighting, tidal states, weather patterns, building a robust training dataset.

Ireland’s plastic problem

The urgency of this work becomes clear when you look at the Marine Institute’s work. Ireland’s 3,172 kilometres of coastline, the longest per capita in Europe, faces a deepening crisis.

A 2018 study found that 73% of deep-sea fish in Irish waters had ingested plastic particles. More than 250 species, including seabirds, fish, marine turtles and mammals have all been reported to ingest large items of plastics.

The costs go beyond harming wildlife, and the economic impact can be significant.

Our drone surveys revealed that some stretches of coast accumulate plastic at rates five to ten times higher than neighbouring areas, driven by ocean currents and river mouths. Without systematic monitoring, these hotspots go unaddressed.

Making the technology accessible

The plastic detection platform accepts drone imagery from any source, such as ordinary people flying their own drones.

Processing requires only standard laptop software. Users upload footage and receive GPS coordinates showing detected plastic locations. The mobile app, available free on iOS and Android, displays these locations as an interactive map.

Community groups, schools and individuals can see nearby plastic pollution and find it, saving a lot of time.

It has already been tested with five community groups around Ireland with positive results, averaging 30 plastics spotted per ten-minute drone flight, varying by location.

Working through the EU-funded BluePoint project, which is tackling plastic pollution of coastlines around Europe, we’ve distributed over 30 drones to partners across Ireland and Europe, including county councils and environmental organisations.

The technology has been deployed in areas including Spanish Point in County Clare, where the local Tidy Towns group (litter-picking volunteers), were named joint Clean Coast Community Group of the Year 2024.

Organising a litter pick. Video by Propeller BIC (Waterford).

The wider waste story

This is part of a broader European effort to address plastic pollution. Partners such as the sports store Decathlon are exploring how to transform recovered beach plastics into new consumer products – sports equipment, textiles and components.

The challenge isn’t just collection. Beach plastics arrive contaminated with sand and salt, in mixed types and grades. Our ongoing research characterises what’s actually found on Irish coastlines, providing manufacturers with data to design appropriate sorting and recycling processes.

The open source software platforms and the drone technology have already been used in nine countries, engaging more than 2,000 people. Pilot programmes are running in France, Spain, Portugal, Brazil and the UK. What began as a question about making beach clean-ups more effective has evolved into a practical system connecting citizen action to environmental outcomes.

Community feedback from pilots has been overwhelmingly positive. Groups report that the drone-derived GPS coordinates transform clean-up work. One participating Tidy Towns group said that volunteers now head straight to flagged locations.

Groups have also reported increased participation, the gamification aspect appeals to families and participants who might not volunteer otherwise. Additionally, the data we’ve gathered so far is being used by local authorities to understand litter patterns and inform policy decisions around waste management and coastal protection.

Gerard Dooly, Assistant Professor in Engineering, University of Limerick

This article is republished from The Conversation under a Creative Commons license. Read the original article.

A system developed by researchers at the University of Waterloo lets people collaborate with groups of robots to create works of art inspired by music.

The new technology features multiple wheeled robots about the size of soccer balls that trail coloured light as they move within a fixed area on the floor in response to key features of music including tempo and chord progression.

A camera records the co-ordinated light trails as they snake within that area, which serves as the canvas for the creation of a “painting,” or visual representation of the emotional content of a particular piece of music.

“Basically, we programmed a swarm of robots to paint based on musical input,” said Dr Gennaro Notomista, a professor of electrical and computer engineering at Waterloo.

“The result is a cohesive system that not only processes musical input, but also co-ordinates multiple painting robots to create adaptive, expressive art that reflects the emotional essence of the music being played.”

The robots represent emotion as they “listen” to music via the colours, intensity and width of their lights trails, as well as their position on the canvas and the speed with which they move within it.

People can simultaneously influence a painting in progress using controls to change the width of light trails and their location on the virtual canvas.

“We included the human control input to allow people and robots to work together,” said Notomista, whose interests include the intersection of art and technology. “The human painter should complement and be complemented by what the robots do.”

The first challenge for researchers was developing an algorithm to control multiple robots within a given area. They tested the system with up to 12 robots, but it can be scaled to handle any number.

Step two involved creating technology to extract and analyze musical features that express emotion so they can then be translated into light trails that appropriately represent them.

Lessons learned during the project have potential applications in other areas requiring the control and co-ordination of multiple robots working in unison, such as environmental monitoring, precision agriculture, search and rescue missions, and planetary exploration.

The research also reflects the University of Waterloo’s Global Futures initiative, which advances interdisciplinary work that considers how emerging technologies can shape society, culture and the human experience.

Later, Notomista plans to enlist professional painters and musicians to explore the possibilities of the new tool in user studies and stage public exhibitions.

A paper on the system, Music-driven Robot Swarm Painting, by Notomista and Jingde Cheng, a former Waterloo graduate student, was presented at the 2025 IEEE International Conference on Advanced Robotics and its Social Impacts.

The dexterity gap: from human hand to robot hand

Observe your own hand. As you read this, it’s holding your phone or clicking your mouse with seemingly effortless grace. With over 20 degrees of freedom, human hands possess extraordinary dexterity, which can grip a heavy hammer, rotate a screwdriver, or instantly adjust when something slips.

With a similar structure to human hands, dexterous robot hands offer great potential:

Universal adaptability: Handling various objects from delicate needles to basketballs, adapting to each unique challenge in real time.

Fine manipulation: Executing complex tasks like key rotation, scissor use, and surgical procedures that are impossible with simple grippers.

Skill transfer: Their similarity to human hands makes them ideal for learning from vast human demonstration data.

Despite this potential, most current robots still rely on simple “grippers” due to the difficulties of dexterous manipulation. The pliers-like grippers are capable only of repetitive tasks in structured environments. This “dexterity gap” severely limits robots’ role in our daily lives.

Among all manipulation skills, grasping stands as the most fundamental. It is the gateway through which many other capabilities emerge. Without reliable grasping, robots cannot pick up tools, manipulate objects, or perform complex tasks. Therefore, we focus on equipping dexterous robots with the capability to robustly grasp diverse objects in this work.

The challenge: why dexterous grasping remains elusive

While humans can grasp almost any object with minimal conscious effort, the path to dexterous robotic grasping is fraught with fundamental challenges that have stymied researchers for decades:

High-dimensional control complexity. With 20+ degrees of freedom, dexterous hands present an astronomically large control space. Each finger’s movement affects the entire grasp, making it extremely difficult to determine optimal finger trajectories and force distributions in real-time. Which finger should move? How much force should be applied? How to adjust in real-time? These seemingly simple questions reveal the extraordinary complexity of dexterous grasping.

Generalization across diverse object shapes. Different objects demand fundamentally different grasp strategies. For example, spherical objects require enveloping grasps, while elongated objects need precision grips. The system must generalize across this vast diversity of shapes, sizes, and materials without explicit programming for each category.

Shape uncertainty under monocular vision. For practical deployment in daily life, robots must rely on single-camera systems—the most accessible and cost-effective sensing solution. Furthermore, we cannot assume prior knowledge of object meshes, CAD models, or detailed 3D information. This creates fundamental uncertainty: depth ambiguity, partial occlusions, and perspective distortions make it challenging to accurately perceive object geometry and plan appropriate grasps.

Our approach: RobustDexGrasp

To address these fundamental challenges, we present RobustDexGrasp, a novel framework that tackles each challenge with targeted solutions:

Teacher-student curriculum for high-dimensional control. We trained our system through a two-stage reinforcement learning process: first, a “teacher” policy learns ideal grasping strategies with privileged information (full object shape and tactile sensors) through extensive exploration in simulation. Then, a “student” policy learns from the teacher using only real-world perception (single-view point cloud, noisy joint positions) and adapts to real-world disturbances.

Hand-centric “intuition” for shape generalization. Instead of capturing complete 3D shape features, our method creates a simple “mental map” that only answers one question: “Where are the surfaces relative to my fingers right now?” This intuitive approach ignores irrelevant details (like color or decorative patterns) and focuses only on what matters for the grasp. It’s the difference between memorizing every detail of a chair versus just knowing where to put your hands to lift it—one is efficient and adaptable, the other is unnecessarily complicated.

Multi-modal perception for uncertainty reduction. Instead of relying on vision alone, we combine the camera’s view with the hand’s “body awareness” (proprioception—knowing where its joints are) and reconstructed “touch sensation” to cross-check and verify what it’s seeing. It’s like how you might squint at something unclear, then reach out to touch it to be sure. This multi-sense approach allows the robot to handle tricky objects that would confuse vision-only systems—grasping a transparent glass becomes possible because the hand “knows” it’s there, even when the camera struggles to see it clearly.

The results: from laboratory to reality

Trained on just 35 simulated objects, our system demonstrates excellent real-world capabilities:

Generalization: It achieved a 94.6% success rate across a diverse test set of 512 real-world objects, including challenging items like thin boxes, heavy tools, transparent bottles, and soft toys.

Robustness: The robot could maintain a secure grip even when a significant external force (equivalent to a 250g weight) was applied to the grasped object, showing far greater resilience than previous state-of-the-art methods.

Adaptation: When objects were accidentally bumped or slipped from its grasp, the policy dynamically adjusted finger positions and forces in real-time to recover, showcasing a level of closed-loop control previously difficult to achieve.

Beyond picking things up: enabling a new era of robotic manipulation

RobustDexGrasp represents a crucial step toward closing the dexterity gap between humans and robots. By enabling robots to grasp nearly any object with human-like reliability, we’re unlocking new possibilities for robotic applications beyond grasping itself. We demonstrated how it can be seamlessly integrated with other AI modules to perform complex, long-horizon manipulation tasks:

Grasping in clutter: Using an object segmentation model to identify the target object, our method enables the hand to pick a specific item from a crowded pile despite interference from other objects.

Task-oriented grasping: With a vision language model as the high-level planner and our method providing the low-level grasping skill, the robot hand can execute grasps for specific tasks, such as cleaning up the table or playing chess with a human.

Dynamic interaction: Using an object tracking module, our method can successfully control the robot hand to grasp objects moving on a conveyor belt.

Looking ahead, we aim to overcome current limitations, such as handling very small objects (which requires a smaller, more anthropomorphic hand) and performing non-prehensile interactions like pushing. The journey to true robotic dexterity is ongoing, and we are excited to be part of it.

Read the work in full

• RobustDexGrasp: Robust Dexterous Grasping of General Objects, Hui Zhang, Zijian Wu, Linyi Huang, Sammy Christen, Jie Song.

By Silvia Cernea Clark

Researchers at Rice University have developed a soft robotic arm capable of performing complex tasks such as navigating around an obstacle or hitting a ball, guided and powered remotely by laser beams without any onboard electronics or wiring. The research could inform new ways to control implantable surgical devices or industrial machines that need to handle delicate objects.

In a proof-of-concept study that integrates smart materials, machine learning and an optical control system, a team of Rice researchers led by materials scientist Hanyu Zhu used a light-patterning device to precisely induce motion in a robotic arm made from azobenzene liquid crystal elastomer ⎯ a type of polymer that responds to light.

According to the study published in Advanced Intelligent Systems, the new robotic system incorporates a neural network trained to predict the exact light pattern needed to create specific arm movements. This makes it easier for the robot to execute complex tasks without needing similarly complex input from an operator.

“This was the first demonstration of real-time, reconfigurable, automated control over a light-responsive material for a soft robotic arm,” said Elizabeth Blackert, a Rice doctoral alumna who is the first author on the study.

Elizabeth Blackert and Hanyu Zhu (Photo credit: Jeff Fitlow/Rice University).

Conventional robots typically involve rigid structures with mobile elements like hinges, wheels or grippers to enable a predefined, relatively constrained range of motion. Soft robots have opened up new areas of application in contexts like medicine, where safely interacting with delicate objects is required. So-called continuum robots are a type of soft robot that forgoes mobility constraints, enabling adaptive motion with a vastly expanded degree of freedom.

“A major challenge in using soft materials for robots is they are either tethered or have very simple, predetermined functionality,” said Zhu, assistant professor of materials science and nanoengineering. “Building remotely and arbitrarily programmable soft robots requires a unique blend of expertise involving materials development, optical system design and machine learning capabilities. Our research team was uniquely suited to take on this interdisciplinary work.”

The team created a new variation of an elastomer that shrinks under blue laser light then relaxes and regrows in the dark ⎯ a feature known as fast relaxation time that makes real-time control possible. Unlike other light-sensitive materials that require harmful ultraviolet light or take minutes to reset, this one works with safer, longer wavelengths and responds within seconds.

“When we shine a laser on one side of the material, the shrinking causes the material to bend in that direction,” Blackert said. “Our material bends toward laser light like a flower stem does toward sunlight.”

To control the material, the researchers used a spatial light modulator to split a single laser beam into multiple beamlets, each directed to a different part of the robotic arm. The beamlets can be turned on or off and adjusted in intensity, allowing the arm to bend or contract at any given point, much like the tentacles of an octopus. This technique can in principle create a robot with virtually infinite degrees of freedom ⎯ far beyond the capabilities of traditional robots with fixed joints.

“What is new here is using the light pattern to achieve complex changes in shape,” said Rafael Verduzco, professor and associate chair of chemical and biomolecular engineering and professor of materials science and nanoengineering. “In prior work, the material itself was patterned or programmed to change shape in one way, but here the material can change in multiple ways, depending on the laser beamlet pattern.”

To train such a multiparameter arm, the team ran a small number of combinations of light settings and recorded how the robot arm deformed in each case, using the data to train a convolutional neural network ⎯ a type of artificial intelligence used in image recognition. The model was then able to output the exact light pattern needed to create a desired shape such as flexing or a reach-around motion.

The current prototype is flat and moves in 2D, but future versions could bend in three dimensions with additional sensors and cameras.

Photo credit: Jeff Fitlow/Rice University

“This is a step towards having safer, more capable robotics for various applications ranging from implantable biomedical devices to industrial robots that handle soft goods,” Blackert said.

For many of us, artificial intelligence (AI) has become part of everyday life, and the rate at which we assign previously human roles to AI systems shows no signs of slowing down. AI systems are the crucial ingredients of many technologies — e.g., self-driving cars, smart urban planning, digital assistants — across a growing number of domains. At the core of many of these technologies are autonomous agents — systems designed to act on behalf of humans and make decisions without direct supervision. In order to act effectively in the real world, these agents must be capable of carrying out a wide range of tasks despite possibly unpredictable environmental conditions, which often requires some form of machine learning (ML) for achieving adaptive behaviour.

Reinforcement learning (RL) [6] stands out as a powerful ML technique for training agents to achieve optimal behaviour in stochastic environments. RL agents learn by interacting with their environment: for every action they take, they receive context-specific rewards or penalties. Over time, they learn behaviour that maximizes the expected rewards throughout their runtime.

Image provided by the authors – generated using Gemini.

RL agents can master a wide variety of complex tasks, from winning video games to controlling cyber-physical systems such as self-driving cars, often surpassing what expert humans are capable of. This optimal, efficient behaviour, however, if left entirely unconstrained, may turn out to be off-putting or even dangerous to the humans it impacts. This motivates the substantial research effort in safe RL, where specialized techniques are developed to ensure that RL agents meet specific safety requirements. These requirements are often expressed in formal languages like linear temporal logic (LTL), which extends classical (true/false) logic with temporal operators, allowing us to specify conditions like “something that must always hold”, or “something that must eventually occur”. By combining the adaptability of ML with the precision of logic, researchers have developed powerful methods for training agents to act both effectively and safely.

However, safety isn’t everything. Indeed, as RL-based agents are increasingly given roles that either replace or closely interact with humans, a new challenge arises: ensuring their behavior is also compliant with the social, legal and ethical norms that structure human society, which often go beyond simple constraints guaranteeing safety. For example, a self-driving car might perfectly follow safety constraints (e.g. avoiding collisions), yet still adopt behaviors that, while technically safe, violate social norms, appearing bizarre or rude on the road, which might cause other (human) drivers to react in unsafe ways.

Norms are typically expressed as obligations (“you must do it”), permissions (“you are permitted to do it”) and prohibitions (“you are forbidden from doing it”), which are not statements that can be true or false, like classical logic formulas. Instead, they are deontic concepts: they describe what is right, wrong, or permissible — ideal or acceptable behaviour, instead of what is actually the case. This nuance introduces several difficult dynamics to reasoning about norms, which many logics (such as LTL) struggle to handle. Even every-day normative systems like driving regulations can feature such complications; while some norms can be very simple (e.g., never exceed 50 kph within city limits), others can be more complex, as in:

1. Always maintain 10 meters between your vehicle and the vehicles in front of and behind you.
2. If there are less than 10 meters between you and the vehicle behind you, you should slow down to put more space between yourself and the vehicle in front of you.

(2) is an example of a contrary-to-duty obligation (CTD), an obligation you must follow specifically in a situation where another primary obligation (1) has already been violated to, e.g., compensate or reduce damage. Although studied extensively in the fields of normative reasoning and deontic logic, such norms can be problematic for many basic safe RL methods based on enforcing LTL constraints, as was discussed in [4].

However, there are approaches for safe RL that show more potential. One notable example is the Restraining Bolt technique, introduced by De Giacomo et al. [2]. Named after a device used in the Star Wars universe to curb the behavior of droids, this method influences an agent’s actions to align with specified rules while still allowing it to pursue its goals. That is, the restraining bolt modifies the behavior an RL agent learns so that it also respects a set of specifications. These specifications, expressed in a variant of LTL (LTLf [3]), are each paired with its own reward. The central idea is simple but powerful: along with the rewards the agent receives while exploring the environment, we add an additional reward whenever its actions satisfy the corresponding specification, nudging it to behave in ways that align with individual safety requirements. The assignment of specific rewards to individual specifications allows us to model more complicated dynamics like, e.g., CTD obligations, by assigning one reward for obeying the primary obligation, and a different reward for obeying the CTD obligation.

Still, issues with modeling norms persist; for example, many (if not most) norms are conditional. Consider the obligation stating “if pedestrians are present at a pedestrian crossing, THEN the nearby vehicles must stop”. If an agent were rewarded every time this rule was satisfied, it would also receive rewards in situations where the norm is not actually in force. This is because, in logic, an implication holds also when the antecedent (“pedestrians are present”) is false. As a result, the agent is rewarded whenever pedestrians are not around, and might learn to prolong its runtime in order to accumulate these rewards for effectively doing nothing, instead of efficiently pursuing its intended task (e.g., reaching a destination). In [5] we showed that there are scenarios where an agent will either ignore the norms, or learn this “procrastination” behavior, no matter which rewards we choose. As a result, we introduced Normative Restraining Bolts (NRBs), a step forward toward enforcing norms in RL agents. Unlike the original Restraining Bolt, which encouraged compliance by providing additional rewards, the normative version instead punishes norm violations. This design is inspired by the Andersonian view of deontic logic [1], which treats obligations as rules whose violation necessarily triggers a sanction. Thus, the framework no longer relies on reinforcing acceptable behavior, but instead enforces norms by guaranteeing that violations carry tangible penalties. While effective for managing intricate normative dynamics like conditional obligations, contrary-to-duties, and exceptions to norms, NRBs rely on trial-and-error reward tuning to implement norm adherence, and therefore can be unwieldy, especially when trying to resolve conflicts between norms. Moreover, they require retraining to accommodate norm updates, and do not lend themselves to guarantees that optimal policies minimize norm violations.

Our contribution

Building on NRBs, we introduce Ordered Normative Restraining Bolts (ONRBs), a framework for guiding reinforcement learning agents to comply with social, legal, and ethical norms while addressing the limitations of NRBs. In this approach, each norm is treated as an objective in a multi-objective reinforcement learning (MORL) problem. Reformulating the problem in this way allows us to:

• Prove that when norms do not conflict, an agent who learns optimal behaviour will minimize norm violations over time.
• Express relationships between norms in terms of a ranking system describing which norm should be prioritized when a conflict occurs.
• Use MORL techniques to algorithmically determine the necessary magnitude of the punishments we assign such that it is guarantied that so long as an agent learns optimal behaviour, norms will be violated as little as possible, prioritizing the norms with the highest rank.
• Accommodate changes in our normative systems by “deactivating” or “reactivating” specific norms.

We tested our framework in a grid-world environment inspired by strategy games, where an agent learns to collect resources and deliver them to designated areas. This setup allows us to demonstrate the framework’s ability to handle the complex normative scenarios we noted above, along with direct prioritization of conflicting norms and norm updates. For instance, the figure below

displays how the agent handles norm conflicts, when it is both obligated to (1) avoid the dangerous (pink) areas, and (2) reach the market (blue) area by a certain deadline, supposing that the second norm takes priority. We can see that it chooses to violate (1) once, because otherwise it will be stuck at the beginning of the map, unable to fulfill (2). Nevertheless, when given the possibility to violate (1) once more, it chooses the compliant path, even though the violating path would allow it to collect more resources, and therefore more rewards from the environment.

In summary, by combining RL with logic, we can build AI agents that do not just work, they work right.

This work won a distinguished paper award at IJCAI 2025. Read the paper in full: Combining MORL with restraining bolts to learn normative behaviour, Emery A. Neufeld, Agata Ciabattoni and Radu Florin Tulcan.

Acknowledgements

This research was funded by the Vienna Science and Technology Fund (WWTF) project ICT22-023 and the Austrian Science Fund (FWF) 10.55776/COE12 Cluster of Excellence Bilateral AI.

References

[1] Alan Ross Anderson. A reduction of deontic logic to alethic modal logic. Mind, 67(265):100–103, 1958.

[2] Giuseppe De Giacomo, Luca Iocchi, Marco Favorito, and Fabio Patrizi. Foundations for restraining bolts: Reinforcement learning with LTLf/LDLf restraining specifications. In Proceedings of the international conference on automated planning and scheduling, volume 29, pages 128–136, 2019.

[3] Giuseppe De Giacomo and Moshe Y Vardi. Linear temporal logic and linear dynamic logic on finite traces. In IJCAI, volume 13, pages 854–860, 2013.

[4] Emery Neufeld, Ezio Bartocci, and Agata Ciabattoni. On normative reinforcement learning via safe reinforcement learning. In PRIMA 2022, 2022.

[5] Emery A Neufeld, Agata Ciabattoni, and Radu Florin Tulcan. Norm compliance in reinforcement learning agents via restraining bolts. In Legal Knowledge and Information Systems JURIX 2024, pages 119–130. IOS Press, 2024.

[6] Richard S. Sutton and Andrew G. Barto. Reinforcement learning – an introduction. Adaptive computation and machine learning. MIT Press, 1998.

By Sean Nealon

Researchers are closer to equipping a dog-like robot to conduct science on the surface of Mars after five days of experiments this month at White Sands National Park in New Mexico.

The national park is serving as a Mars analog environment and the scientists are conducting field test scenarios to inform future Mars operations with astronauts, dog-like robots known as quadruped robots, rovers and scientists at Mission Control on Earth. The work builds on similar experiments by the team with the same robot on the slopes of Mount Hood in Oregon, which simulated the landscape on the Moon.

“Our group is very committed to putting quadrupeds on the Moon and on Mars,” said Cristina Wilson, a robotics researcher in the College of Engineering at Oregon State University. “It’s the next frontier and takes advantage of the unique capabilities of legged robots.”

The NASA-funded project supports the agency’s Moon to Mars program, which is developing the tools for long-term lunar exploration and future crewed missions to Mars. It builds on research that has enabled NASA to send rovers and a helicopter to Mars.

The LASSIE Project: Legged Autonomous Surface Science in Analog Environments includes engineers, cognitive scientists, geoscientists and planetary scientists from Oregon State, the University of Southern California, Texas A&M University, the Georgia Institute of Technology, the University of Pennsylvania, Temple University and NASA Johnson Space Center.

The field work this month at White Sands was the second time the research team visited the national park. They made the initial trip in 2023 and also made trips in 2023 and 2024 to Mount Hood. During these field sessions, the scientists gather data from the feet of the quadruped robots, which can measure mechanical responses to foot-surface interactions.

“In the same way that the human foot standing on ground can sense the stability of the surface as things shift, legged robots are capable of potentially feeling the exact same thing,” Wilson said. “So each step the robot takes provides us information that will help its future performance in places like the Moon or Mars.”

Quadruped robot.

The conditions at White Sands this month were challenging. Triple-digit high temperatures meant the team started field work at sunrise and wrapped by late morning because of the rising heat index and its impact on the researchers and the power supply to the robots.

But the team made important progress. Improvements to the algorithms they have refined in recent years led for the first time to the robot acting autonomously and making its own decisions.

This is important, Wilson noted, because in a scenario where the quadruped would be on the surface of Mars with an astronaut, it would allow both the robot and the astronaut to act independently, increasing the amount of scientific work that could be accomplished.

They also tested advances they have made in developing different ways for the robot to move depending on surface conditions, which could lead to increased energy efficiency, Wilson said.

“There is certainly a lot more research to do, but these are important steps in realizing the goal of sending quadrupeds to the Moon and Mars,” Wilson said.

Other leaders of the project include Feifei Qian, USC; Ryan Ewing and Kenton Fisher, NASA Johnson Space Center; Marion Nachon, Texas A&M; Frances Rivera-Hernández, Georgia Tech; Douglas Jerolmack and Daniel Koditschek, University of Pennsylvania; and Thomas Shipley, Temple University.

The research is funded by the NASA Planetary Science and Technology through Analog Research (PSTAR) program, and Mars Exploration Program.

Thanks to those that sent and suggested AI and robotics-themed holiday videos, images, and stories. Here’s a sample to get you into the spirit this season.

Josh Miller used ChatGPT to create the 12 days of Christmas. Prompt: “I want to make 12 images in an early 1980s era to depict the twelve days of Christmas”.

See the pictures in higher definition here.

“Santa’s Secret: Robots Saved Christmas! | CPS XMAS video 2023” from Chair of Cyber-Physical Systems.

“AI rewriting classic Christmas stories” from GingerPale.

Using PaLM2 to generate haikus.

Prompt: “Write me a haiku about an alternative xmas”.
No gifts, no grand feast,
Just laughter and joy, warmth spread,
Kindness fills the air.

Prompt: “Write a haiku about the winter holiday season”.
Snow blankets the land,
Hearts aglow with love’s warmth,
Joyous cheer surrounds.

“Put a Bow on It | Happy Holidays” from Boston Dynamics.

“Christmas Magic: Animated Backgrounds for an Enchanting Atmosphere” from AI-ArtificialIntelligence.

“XMAS – Episode 3: AI – Christmas music for the HAKOM Christmas party” from HAKOM Time Series.

“Christmas tree in python notebook with Turtle” from Learn Upwards.

“FZI Living Lab Christmas Robotics 2023 – Robotic Reindeer Rescue” from FZI channel.

A letter to Santa generated by PaLM2. Prompt: “Write me a letter to Santa Claus. I don’t want any presents, just inspiration in my machine learning research for 2024.”

Dear Santa Claus,

As the year comes to a close, I find myself reflecting on the remarkable advancements in machine learning that have transformed our world. From self-driving cars to personalized medicine, these innovations have opened up new possibilities and enriched our lives in ways we could have only imagined a few decades ago.

I am writing to you today not to ask for material gifts, but rather for something more precious: inspiration to continue exploring the frontiers of machine learning in the year to come. As a researcher in this field, I am eager to contribute to the development of algorithms and techniques that can further enhance our understanding of the world around us and create solutions to pressing global challenges.

I am particularly fascinated by the potential of machine learning to address the climate crisis. By harnessing the power of data and computation, we can develop more efficient renewable energy sources, optimize transportation systems, and mitigate the impact of extreme weather events. Additionally, I believe that machine learning can play a crucial role in improving education, healthcare, and environmental sustainability.

I know that you have a special ability to see the world in a unique and insightful way. I hope that you can share some of your wisdom with me and guide me towards promising research directions. I am eager to embark on new challenges and make a positive impact on the world through machine learning.

Thank you for considering my request, Santa Claus. I look forward to your guidance and inspiration as I pursue my research endeavors in the year ahead.

With warmest regards,

[Your Name]

Be it the news or the dwindling number of creatures hitting your windscreens, it will not have evaded you that the insect world in bad shape.

In the last three decades, the global biomass of flying insects has shrunk by 75%. Among the trend’s most notables victims is the world’s most important pollinator, the honeybee. In the United States, 48% of honeybee colonies died in 2023 alone, making it the second deadliest year on record. This significant loss is due in part to colony collapse disorder (CCD), the sudden disappearance of bees. In contrast, European countries report lower but still worrisome rates of colony losses, ranging from 6% to 32%.

This decline causes many of our essential food crops to be under-pollinated, a phenomenon that threatens our society’s food security.

Debunking the sci-fi myth of robotic bees

So, what can be done? Given pesticides’ role in the decline of bee colonies, commonly proposed solutions include a shift away from industrial farming and toward less pesticide-intensive, more sustainable forms of agriculture.

Others tend to look toward the sci-fi end of things, with some scientists imagining that we could eventually replace live honeybees with robotic ones. Such artificial bees could interact with flowers like natural insects, maintaining pollination levels despite the declining numbers of natural pollinators. The vision of artificial pollinators contributed to ingenious designs of insect-sized robots capable of flying.

In reality, such inventions are more effective at educating us over engineers’ fantasies than they are at reviving bee colonies, so slim are their prospects of materialising. First, these artificial pollinators would have to be equipped for much more more than just flying. Daily tasks carried out by the common bee include searching for plants, identifying flowers, unobtrusively interacting with them, locating energy sources, ducking potential predators, and dealing with adverse weather conditions. Robots would have to perform all of these in the wild with a very high degree of reliability since any broken-down or lost robot can cause damage and spread pollution. Second, it remains to be seen whether our technological knowledge would be even capable of manufacturing such inventions. This is without even mentioning the price tag of a swarm of robots capable of substituting pollination provided by a single honeybee colony.

Inside a smart hive

Rather than trying to replace honeybees with robots, our two latest projects funded by the European Union propose that the robots and honeybees actually team up. Were these to succeed, struggling honeybee colonies could be transformed into bio-hybrid entities consisting of biological and technological components with complementary skills. This would hopefully boost and secure the colonies’ population growth as more bees survive over harsh winters and yield more foragers to pollinate surrounding ecosystems.

The first of these projects, Hiveopolis, investigates how the complex decentralised decision-making mechanism in a honeybee colony can be nudged by digital technology. Begun in 2019 and set to end in March 2024, the experiment introduces technology into three observation hives each containing 4,000 bees, by contrast to 40,000 bees for a normal colony.

Within this honeybee smart home, combs have integrated temperature sensors and heating devices, allowing the bees to enjoy optimal conditions inside the colony. Since bees tend to snuggle up to warmer locations, the combs also enables us to direct them toward different areas of the hive. And as if that control weren’t enough, the hives are also equipped with a system of electronic gates that monitors the insects movements. Both technologies allow us to decide where the bees store honey and pollen, but also when they vacate the combs so as to enable us to harvest honey. Last but not least, the smart hive contains a robotic dancing bee that can direct foraging bees toward areas with plants to be pollinated.

Due to the experiment’s small scale, it is impossible to draw conclusions on the extent to which our technologies may have prevented bee losses. However, there is little doubt what we have seen thus far give reasons to be hopeful. We can confidently assert that our smart beehives allowed colonies to survive extreme cold during the winter in a way that wouldn’t otherwise be possible. To precisely assess how many bees these technologies have saved would require upscaling the experiment to hundreds of colonies.

Pampering the queen bee

Our second EU-funded project, RoboRoyale, focuses on the honeybee queen and her courtyard bees, with robots in this instance continuously monitoring and interacting with her Royal Highness.

Come 2024, we will equip each hive with a group of six bee-sized robots, which will groom and feed the honeybee queen to affect the number of eggs she lays. Some of these robots will be equipped with royal jelly micro-pumps to feed her, while others will feature compliant micro-actuators to groom her. These robots will then be connected to a larger robotic arm with infrared cameras, that will continuously monitor the queen and her vicinity.

As witnessed by the photo to the right and also below, we have already been able to successfully introduce the robotic arm within a living colony. There it continuously monitored the queen and determined her whereabouts through light stimuli.

Emulating the worker bees

In a second phase, it is hoped the bee-sized robots and robotic arm will be able to emulate the behaviour of the workers, the female bees lacking reproductive capacity who attend to the queen and feed her royal jelly. Rich in water, proteins, carbohydrates, lipids, vitamins and minerals, this nutritious substance secreted by the glands of the worker bees enables the queen to lay up to thousands of eggs a day.

Worker bees also engage in cleaning the queen, which involves licking her. During such interactions, they collect some of the queen’s pheromones and disperse them throughout the colony as they move across the hive. The presence of these pheromones controls many of the colony’s behaviours and notifies the colony of a queen’s presence. For example, in the event of the queen’s demise, a new queen must be quickly reared from an egg laid by the late queen, leaving only a narrow time window for the colony to react.

Finally, it is believed worker bees may also act as the queen’s guides, leading her to laying eggs in specific comb cells. The size of these cells can determine if the queen lays a diploid or haploid egg, resulting in the bee developing into either into drone (male) or worker (female) bee. Taking over these guiding duties could affect no less than the rate’s entire reproductive rate.

How robots can prevent bee cannibalism

This could have another virtuous effect: preventing cannibalism.

During tough times, such as long periods of rain, bees have to make do with little pollen intake. This forces them to feed young larvae to older ones so that at least the older larvae has a chance to survive. Through RoboRoyale, we will look not only to reduce chances of this behaviour occurring, but also quantify to what extent it occurs under normal conditions.

Ultimately, our robots will enable us to deepen our understanding of the very complex regulation processes inside honeybee colonies through novel experimental procedures. The insights gained from these new research tracks will be necessary to better protect these valuable social insects and ensure sufficient pollination in the future – a high stakes enterprise for food security.

This article is the result of The Conversation’s collaboration with Horizon, the EU research and innovation magazine.

Farshad Arvin is a member of the Department of Computer Science at Durham University in the UK. The research of Farshad Arvin is primarily funded by the EU H2020 and Horizon Europe programmes.

Martin Stefanec is a member of the Institute of Biology at the University of Graz. He has received funding from the EU programs H2020 and Horizon Europe.

Tomas Krajnik is member of the Institute of Electrical and Electronics Engineers (IEEE). The research of Tomas Krajnik is primarily funded by EU H2020 Horizon programme and Czech National Science Foundation.

Feature Fields for Robotic Manipulation (F3RM) enables robots to interpret open-ended text prompts using natural language, helping the machines manipulate unfamiliar objects. The system’s 3D feature fields could be helpful in environments that contain thousands of objects, such as warehouses. Images courtesy of the researchers.

By Alex Shipps | MIT CSAIL

Imagine you’re visiting a friend abroad, and you look inside their fridge to see what would make for a great breakfast. Many of the items initially appear foreign to you, with each one encased in unfamiliar packaging and containers. Despite these visual distinctions, you begin to understand what each one is used for and pick them up as needed.

Inspired by humans’ ability to handle unfamiliar objects, a group from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) designed Feature Fields for Robotic Manipulation (F3RM), a system that blends 2D images with foundation model features into 3D scenes to help robots identify and grasp nearby items. F3RM can interpret open-ended language prompts from humans, making the method helpful in real-world environments that contain thousands of objects, like warehouses and households.

F3RM offers robots the ability to interpret open-ended text prompts using natural language, helping the machines manipulate objects. As a result, the machines can understand less-specific requests from humans and still complete the desired task. For example, if a user asks the robot to “pick up a tall mug,” the robot can locate and grab the item that best fits that description.

“Making robots that can actually generalize in the real world is incredibly hard,” says Ge Yang, postdoc at the National Science Foundation AI Institute for Artificial Intelligence and Fundamental Interactions and MIT CSAIL. “We really want to figure out how to do that, so with this project, we try to push for an aggressive level of generalization, from just three or four objects to anything we find in MIT’s Stata Center. We wanted to learn how to make robots as flexible as ourselves, since we can grasp and place objects even though we’ve never seen them before.”

Learning “what’s where by looking”

The method could assist robots with picking items in large fulfillment centers with inevitable clutter and unpredictability. In these warehouses, robots are often given a description of the inventory that they’re required to identify. The robots must match the text provided to an object, regardless of variations in packaging, so that customers’ orders are shipped correctly.

For example, the fulfillment centers of major online retailers can contain millions of items, many of which a robot will have never encountered before. To operate at such a scale, robots need to understand the geometry and semantics of different items, with some being in tight spaces. With F3RM’s advanced spatial and semantic perception abilities, a robot could become more effective at locating an object, placing it in a bin, and then sending it along for packaging. Ultimately, this would help factory workers ship customers’ orders more efficiently.

“One thing that often surprises people with F3RM is that the same system also works on a room and building scale, and can be used to build simulation environments for robot learning and large maps,” says Yang. “But before we scale up this work further, we want to first make this system work really fast. This way, we can use this type of representation for more dynamic robotic control tasks, hopefully in real-time, so that robots that handle more dynamic tasks can use it for perception.”

The MIT team notes that F3RM’s ability to understand different scenes could make it useful in urban and household environments. For example, the approach could help personalized robots identify and pick up specific items. The system aids robots in grasping their surroundings — both physically and perceptively.

“Visual perception was defined by David Marr as the problem of knowing ‘what is where by looking,’” says senior author Phillip Isola, MIT associate professor of electrical engineering and computer science and CSAIL principal investigator. “Recent foundation models have gotten really good at knowing what they are looking at; they can recognize thousands of object categories and provide detailed text descriptions of images. At the same time, radiance fields have gotten really good at representing where stuff is in a scene. The combination of these two approaches can create a representation of what is where in 3D, and what our work shows is that this combination is especially useful for robotic tasks, which require manipulating objects in 3D.”

Creating a “digital twin”

F3RM begins to understand its surroundings by taking pictures on a selfie stick. The mounted camera snaps 50 images at different poses, enabling it to build a neural radiance field (NeRF), a deep learning method that takes 2D images to construct a 3D scene. This collage of RGB photos creates a “digital twin” of its surroundings in the form of a 360-degree representation of what’s nearby.

In addition to a highly detailed neural radiance field, F3RM also builds a feature field to augment geometry with semantic information. The system uses CLIP, a vision foundation model trained on hundreds of millions of images to efficiently learn visual concepts. By reconstructing the 2D CLIP features for the images taken by the selfie stick, F3RM effectively lifts the 2D features into a 3D representation.

Keeping things open-ended

After receiving a few demonstrations, the robot applies what it knows about geometry and semantics to grasp objects it has never encountered before. Once a user submits a text query, the robot searches through the space of possible grasps to identify those most likely to succeed in picking up the object requested by the user. Each potential option is scored based on its relevance to the prompt, similarity to the demonstrations the robot has been trained on, and if it causes any collisions. The highest-scored grasp is then chosen and executed.

To demonstrate the system’s ability to interpret open-ended requests from humans, the researchers prompted the robot to pick up Baymax, a character from Disney’s “Big Hero 6.” While F3RM had never been directly trained to pick up a toy of the cartoon superhero, the robot used its spatial awareness and vision-language features from the foundation models to decide which object to grasp and how to pick it up.

F3RM also enables users to specify which object they want the robot to handle at different levels of linguistic detail. For example, if there is a metal mug and a glass mug, the user can ask the robot for the “glass mug.” If the bot sees two glass mugs and one of them is filled with coffee and the other with juice, the user can ask for the “glass mug with coffee.” The foundation model features embedded within the feature field enable this level of open-ended understanding.

“If I showed a person how to pick up a mug by the lip, they could easily transfer that knowledge to pick up objects with similar geometries such as bowls, measuring beakers, or even rolls of tape. For robots, achieving this level of adaptability has been quite challenging,” says MIT PhD student, CSAIL affiliate, and co-lead author William Shen. “F3RM combines geometric understanding with semantics from foundation models trained on internet-scale data to enable this level of aggressive generalization from just a small number of demonstrations.”

Shen and Yang wrote the paper under the supervision of Isola, with MIT professor and CSAIL principal investigator Leslie Pack Kaelbling and undergraduate students Alan Yu and Jansen Wong as co-authors. The team was supported, in part, by Amazon.com Services, the National Science Foundation, the Air Force Office of Scientific Research, the Office of Naval Research’s Multidisciplinary University Initiative, the Army Research Office, the MIT-IBM Watson Lab, and the MIT Quest for Intelligence. Their work will be presented at the 2023 Conference on Robot Learning.

• PAPER – Distilled Feature Fields Enable Few-Shot Language-Guided Manipulation.
  William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. arXiv

The AAAI Fall Symposium Series took place in Arlington, USA, and comprised seven different symposia. One of these, the tenth Artificial Intelligence for Human-Robot Interaction (AI-HRI) symposium was run as a hybrid in-person/online event, and we tuned in to the opening keynote, which was given by Patrícia Alves-Oliveira.

As a psychology student, Patrícia’s dream was to become a therapist. However, an internship, where she encountered a robot for the first time, inspired her to change her plans, and she decided to go into the field of human-robot interaction. Following a PhD in the field, she worked as a postdoc, before heading to industry as a designer in the Amazon Astro robot team.

Patrícia has worked on a number of interesting projects during her time in academia and in industry. Thinking about how to design robots for specific user needs, and keeping the user at the forefront during the design process, has been core to her work. She began by summarising three very different academic projects.

Creativity and robotics

The objective of this project was to design, fabricate, and evaluate robots as creativity-provoking tools for kids. Patrícia created a social robot named YOLO (or Your Own Living Object) that she designed to be child-proof (in other words, it could withstand being dropped and knocked over), with the aim of trying to help children explore their creativity during play. A machine learning algorithm learns the pattern of play that the child has and adapts the robot behaviour accordingly. You can see the robot in action in the demo below:

FLEXI robot

As a postdoc project, Patrícia worked on building FLEXI, a social robot embodiment kit. This kit consists of a robot (with a face, and a torso with a screen on the front), which can be customised, and an open-source end-user programming interface designed to be user-friendly. The customisation element means that it can be used for many applications. The team has deployed FLEXI across three application scenarios: community-support, mental health, and education, with the aim of assessing the flexibility of the system. You can see the robot in action, in different scenarios, here.

Social dining

This project centred on a robotic arm for people with impaired mobility. Such systems already exist for assisting people with tasks such as eating. However, in a social context they can often form a barrier between the user and the rest of the group. The idea behind this project was to consider how such a robot could be adapted to work well in a social context, for example, during a meal with family or friends. The team interviewed people with impaired mobility to assess their needs, and came up with a set of design principles for creating robot-assisted feeding systems and an implementation guide for future research in this area. You can read the research paper on this project here.

You can find out more about these three projects, and the other projects that Patrícia has been involved in, here.

Astro robot

Patrícia has long been interested in robots for the real world, and how this real-world experience is aligned with the study of robots in academia and industry. She decided to leave academia and join the Astro robot programme, which she felt was a great opportunity to work on a large-scale real-world robot project.

The Astro robot is a home robot designed to assist with tasks such as monitoring your house, delivering small objects within the home, recognising your pet, telling a story, or playing games.

Patrícia took us through a typical day in the life of a designer where she always has in mind the bigger picture of what the team is aiming for, in other words, what the ideal robot, and its interactions with humans, would look like. Coupled to that, the process is governed by core design tenets, such as the customer needs, and non-negotiable core elements that the robot should include. When considering a particular element of the robot design, for example, the delivery of an item in the robot tray, Patrícia uses storyboards to map out details of potential human-robot interactions. An important aspect of design concerns edge cases, which occur regularly in the real world. Good design will consider potential edge cases and incorporate ways to deal with them.

Patrícia closed by emphasising the importance of teamwork in the design process, in particular, the need for interdisciplinary teams; by considering design from many different points of view, the chance of innovation is higher.

You can find out more about the Artificial Intelligence for Human-Robot Interaction (AI-HRI) symposium here.

Image/Shutterstock.com

By Angharad Brewer Gillham, Frontiers science writer

‘Social loafing’ is a phenomenon which happens when members of a team start to put less effort in because they know others will cover for them. Scientists investigating whether this happens in teams which combine work by robots and humans found that humans carrying out quality assurance tasks spotted fewer errors when they had been told that robots had already checked a piece, suggesting they relied on the robots and paid less attention to the work.

Now that improvements in technology mean that some robots work alongside humans, there is evidence that those humans have learned to see them as team-mates — and teamwork can have negative as well as positive effects on people’s performance. People sometimes relax, letting their colleagues do the work instead. This is called ‘social loafing’, and it’s common where people know their contribution won’t be noticed or they’ve acclimatized to another team member’s high performance. Scientists at the Technical University of Berlin investigated whether humans social loaf when they work with robots.

“Teamwork is a mixed blessing,” said Dietlind Helene Cymek, first author of the study in Frontiers in Robotics and AI. “Working together can motivate people to perform well but it can also lead to a loss of motivation because the individual contribution is not as visible. We were interested in whether we could also find such motivational effects when the team partner is a robot.”

A helping hand

The scientists tested their hypothesis using a simulated industrial defect-inspection task: looking at circuit boards for errors. The scientists provided images of circuit boards to 42 participants. The circuit boards were blurred, and the sharpened images could only be viewed by holding a mouse tool over them. This allowed the scientists to track participants’ inspection of the board.

Half of the participants were told that they were working on circuit boards that had been inspected by a robot called Panda. Although these participants did not work directly with Panda, they had seen the robot and could hear it while they worked. After examining the boards for errors and marking them, all participants were asked to rate their own effort, how responsible for the task they felt, and how they performed.

Looking but not seeing

At first sight, it looked as if the presence of Panda had made no difference — there was no statistically significant difference between the groups in terms of time spent inspecting the circuit boards and the area searched. Participants in both groups rated their feelings of responsibility for the task, effort expended, and performance similarly.

But when the scientists looked more closely at participants’ error rates, they realized that the participants working with Panda were catching fewer defects later in the task, when they’d already seen that Panda had successfully flagged many errors. This could reflect a ‘looking but not seeing’ effect, where people get used to relying on something and engage with it less mentally. Although the participants thought they were paying an equivalent amount of attention, subconsciously they assumed that Panda hadn’t missed any defects.

“It is easy to track where a person is looking, but much harder to tell whether that visual information is being sufficiently processed at a mental level,” said Dr Linda Onnasch, senior author of the study.

The experimental set-up with the human-robot team. Image supplied by the authors.

Safety at risk?

The authors warned that this could have safety implications. “In our experiment, the subjects worked on the task for about 90 minutes, and we already found that fewer quality errors were detected when they worked in a team,” said Onnasch. “In longer shifts, when tasks are routine and the working environment offers little performance monitoring and feedback, the loss of motivation tends to be much greater. In manufacturing in general, but especially in safety-related areas where double checking is common, this can have a negative impact on work outcomes.”

The scientists pointed out that their test has some limitations. While participants were told they were in a team with the robot and shown its work, they did not work directly with Panda. Additionally, social loafing is hard to simulate in the laboratory because participants know they are being watched.

“The main limitation is the laboratory setting,” Cymek explained. “To find out how big the problem of loss of motivation is in human-robot interaction, we need to go into the field and test our assumptions in real work environments, with skilled workers who routinely do their work in teams with robots.”

• PAPER – Lean back or lean in? Exploring social loafing in human–robot teams. Dietlind Helene Cymek, Anna Truckenbrodt, and Linda Onnasch. Frontiers in Robotics and AI, 10, 1249252.

By Andre He, Vivek Myers

A longstanding goal of the field of robot learning has been to create generalist agents that can perform tasks for humans. Natural language has the potential to be an easy-to-use interface for humans to specify arbitrary tasks, but it is difficult to train robots to follow language instructions. Approaches like language-conditioned behavioral cloning (LCBC) train policies to directly imitate expert actions conditioned on language, but require humans to annotate all training trajectories and generalize poorly across scenes and behaviors. Meanwhile, recent goal-conditioned approaches perform much better at general manipulation tasks, but do not enable easy task specification for human operators. How can we reconcile the ease of specifying tasks through LCBC-like approaches with the performance improvements of goal-conditioned learning?

Conceptually, an instruction-following robot requires two capabilities. It needs to ground the language instruction in the physical environment, and then be able to carry out a sequence of actions to complete the intended task. These capabilities do not need to be learned end-to-end from human-annotated trajectories alone, but can instead be learned separately from the appropriate data sources. Vision-language data from non-robot sources can help learn language grounding with generalization to diverse instructions and visual scenes. Meanwhile, unlabeled robot trajectories can be used to train a robot to reach specific goal states, even when they are not associated with language instructions.

Conditioning on visual goals (i.e. goal images) provides complementary benefits for policy learning. As a form of task specification, goals are desirable for scaling because they can be freely generated hindsight relabeling (any state reached along a trajectory can be a goal). This allows policies to be trained via goal-conditioned behavioral cloning (GCBC) on large amounts of unannotated and unstructured trajectory data, including data collected autonomously by the robot itself. Goals are also easier to ground since, as images, they can be directly compared pixel-by-pixel with other states.

However, goals are less intuitive for human users than natural language. In most cases, it is easier for a user to describe the task they want performed than it is to provide a goal image, which would likely require performing the task anyways to generate the image. By exposing a language interface for goal-conditioned policies, we can combine the strengths of both goal- and language- task specification to enable generalist robots that can be easily commanded. Our method, discussed below, exposes such an interface to generalize to diverse instructions and scenes using vision-language data, and improve its physical skills by digesting large unstructured robot datasets.

Goal representations for instruction following

The GRIF model consists of a language encoder, a goal encoder, and a policy network. The encoders respectively map language instructions and goal images into a shared task representation space, which conditions the policy network when predicting actions. The model can effectively be conditioned on either language instructions or goal images to predict actions, but we are primarily using goal-conditioned training as a way to improve the language-conditioned use case.

Our approach, Goal Representations for Instruction Following (GRIF), jointly trains a language- and a goal- conditioned policy with aligned task representations. Our key insight is that these representations, aligned across language and goal modalities, enable us to effectively combine the benefits of goal-conditioned learning with a language-conditioned policy. The learned policies are then able to generalize across language and scenes after training on mostly unlabeled demonstration data.

We trained GRIF on a version of the Bridge-v2 dataset containing 7k labeled demonstration trajectories and 47k unlabeled ones within a kitchen manipulation setting. Since all the trajectories in this dataset had to be manually annotated by humans, being able to directly use the 47k trajectories without annotation significantly improves efficiency.

To learn from both types of data, GRIF is trained jointly with language-conditioned behavioral cloning (LCBC) and goal-conditioned behavioral cloning (GCBC). The labeled dataset contains both language and goal task specifications, so we use it to supervise both the language- and goal-conditioned predictions (i.e. LCBC and GCBC). The unlabeled dataset contains only goals and is used for GCBC. The difference between LCBC and GCBC is just a matter of selecting the task representation from the corresponding encoder, which is passed into a shared policy network to predict actions.

By sharing the policy network, we can expect some improvement from using the unlabeled dataset for goal-conditioned training. However,GRIF enables much stronger transfer between the two modalities by recognizing that some language instructions and goal images specify the same behavior. In particular, we exploit this structure by requiring that language- and goal- representations be similar for the same semantic task. Assuming this structure holds, unlabeled data can also benefit the language-conditioned policy since the goal representation approximates that of the missing instruction.

Alignment through contrastive learning

We explicitly align representations between goal-conditioned and language-conditioned tasks on the labeled dataset through contrastive learning.

Since language often describes relative change, we choose to align representations of state-goal pairs with the language instruction (as opposed to just goal with language). Empirically, this also makes the representations easier to learn since they can omit most information in the images and focus on the change from state to goal.

We learn this alignment structure through an infoNCE objective on instructions and images from the labeled dataset. We train dual image and text encoders by doing contrastive learning on matching pairs of language and goal representations. The objective encourages high similarity between representations of the same task and low similarity for others, where the negative examples are sampled from other trajectories.

When using naive negative sampling (uniform from the rest of the dataset), the learned representations often ignored the actual task and simply aligned instructions and goals that referred to the same scenes. To use the policy in the real world, it is not very useful to associate language with a scene; rather we need it to disambiguate between different tasks in the same scene. Thus, we use a hard negative sampling strategy, where up to half the negatives are sampled from different trajectories in the same scene.

Naturally, this contrastive learning setup teases at pre-trained vision-language models like CLIP. They demonstrate effective zero-shot and few-shot generalization capability for vision-language tasks, and offer a way to incorporate knowledge from internet-scale pre-training. However, most vision-language models are designed for aligning a single static image with its caption without the ability to understand changes in the environment, and they perform poorly when having to pay attention to a single object in cluttered scenes.

To address these issues, we devise a mechanism to accommodate and fine-tune CLIP for aligning task representations. We modify the CLIP architecture so that it can operate on a pair of images combined with early fusion (stacked channel-wise). This turns out to be a capable initialization for encoding pairs of state and goal images, and one which is particularly good at preserving the pre-training benefits from CLIP.

Robot policy results

For our main result, we evaluate the GRIF policy in the real world on 15 tasks across 3 scenes. The instructions are chosen to be a mix of ones that are well-represented in the training data and novel ones that require some degree of compositional generalization. One of the scenes also features an unseen combination of objects.

We compare GRIF against plain LCBC and stronger baselines inspired by prior work like LangLfP and BC-Z. LLfP corresponds to jointly training with LCBC and GCBC. BC-Z is an adaptation of the namesake method to our setting, where we train on LCBC, GCBC, and a simple alignment term. It optimizes the cosine distance loss between the task representations and does not use image-language pre-training.

The policies were susceptible to two main failure modes. They can fail to understand the language instruction, which results in them attempting another task or performing no useful actions at all. When language grounding is not robust, policies might even start an unintended task after having done the right task, since the original instruction is out of context.

Examples of grounding failures

“put the mushroom in the metal pot”

“put the spoon on the towel”

“put the yellow bell pepper on the cloth”

“put the yellow bell pepper on the cloth”

The other failure mode is failing to manipulate objects. This can be due to missing a grasp, moving imprecisely, or releasing objects at the incorrect time. We note that these are not inherent shortcomings of the robot setup, as a GCBC policy trained on the entire dataset can consistently succeed in manipulation. Rather, this failure mode generally indicates an ineffectiveness in leveraging goal-conditioned data.

Examples of manipulation failures

“move the bell pepper to the left of the table”

“put the bell pepper in the pan”

“move the towel next to the microwave”

Comparing the baselines, they each suffered from these two failure modes to different extents. LCBC relies solely on the small labeled trajectory dataset, and its poor manipulation capability prevents it from completing any tasks. LLfP jointly trains the policy on labeled and unlabeled data and shows significantly improved manipulation capability from LCBC. It achieves reasonable success rates for common instructions, but fails to ground more complex instructions. BC-Z’s alignment strategy also improves manipulation capability, likely because alignment improves the transfer between modalities. However, without external vision-language data sources, it still struggles to generalize to new instructions.

GRIF shows the best generalization while also having strong manipulation capabilities. It is able to ground the language instructions and carry out the task even when many distinct tasks are possible in the scene. We show some rollouts and the corresponding instructions below.

Policy Rollouts from GRIF

“move the pan to the front”

“put the bell pepper in the pan”

“put the knife on the purple cloth”

“put the spoon on the towel”

Conclusion

GRIF enables a robot to utilize large amounts of unlabeled trajectory data to learn goal-conditioned policies, while providing a “language interface” to these policies via aligned language-goal task representations. In contrast to prior language-image alignment methods, our representations align changes in state to language, which we show leads to significant improvements over standard CLIP-style image-language alignment objectives. Our experiments demonstrate that our approach can effectively leverage unlabeled robotic trajectories, with large improvements in performance over baselines and methods that only use the language-annotated data

Our method has a number of limitations that could be addressed in future work. GRIF is not well-suited for tasks where instructions say more about how to do the task than what to do (e.g., “pour the water slowly”)—such qualitative instructions might require other types of alignment losses that consider the intermediate steps of task execution. GRIF also assumes that all language grounding comes from the portion of our dataset that is fully annotated or a pre-trained VLM. An exciting direction for future work would be to extend our alignment loss to utilize human video data to learn rich semantics from Internet-scale data. Such an approach could then use this data to improve grounding on language outside the robot dataset and enable broadly generalizable robot policies that can follow user instructions.

This post is based on the following paper:

• Goal Representations for Instruction Following: A Semi-Supervised Language Interface to Control
  Vivek Myers*, Andre He*, Kuan Fang, Homer Walke, Philippe Hansen-Estruch, Ching-An Cheng, Mihai Jalobeanu, Andrey Kolobov, Anca Dragan, and Sergey Levine

MIT researchers are using generative AI models to help robots more efficiently solve complex object manipulation problems, such as packing a box with different objects. Image: courtesy of the researchers.

By Adam Zewe | MIT News

Anyone who has ever tried to pack a family-sized amount of luggage into a sedan-sized trunk knows this is a hard problem. Robots struggle with dense packing tasks, too.

For the robot, solving the packing problem involves satisfying many constraints, such as stacking luggage so suitcases don’t topple out of the trunk, heavy objects aren’t placed on top of lighter ones, and collisions between the robotic arm and the car’s bumper are avoided.

Some traditional methods tackle this problem sequentially, guessing a partial solution that meets one constraint at a time and then checking to see if any other constraints were violated. With a long sequence of actions to take, and a pile of luggage to pack, this process can be impractically time consuming.   

MIT researchers used a form of generative AI, called a diffusion model, to solve this problem more efficiently. Their method uses a collection of machine-learning models, each of which is trained to represent one specific type of constraint. These models are combined to generate global solutions to the packing problem, taking into account all constraints at once.

Their method was able to generate effective solutions faster than other techniques, and it produced a greater number of successful solutions in the same amount of time. Importantly, their technique was also able to solve problems with novel combinations of constraints and larger numbers of objects, that the models did not see during training.

Due to this generalizability, their technique can be used to teach robots how to understand and meet the overall constraints of packing problems, such as the importance of avoiding collisions or a desire for one object to be next to another object. Robots trained in this way could be applied to a wide array of complex tasks in diverse environments, from order fulfillment in a warehouse to organizing a bookshelf in someone’s home.

“My vision is to push robots to do more complicated tasks that have many geometric constraints and more continuous decisions that need to be made — these are the kinds of problems service robots face in our unstructured and diverse human environments. With the powerful tool of compositional diffusion models, we can now solve these more complex problems and get great generalization results,” says Zhutian Yang, an electrical engineering and computer science graduate student and lead author of a paper on this new machine-learning technique.

Her co-authors include MIT graduate students Jiayuan Mao and Yilun Du; Jiajun Wu, an assistant professor of computer science at Stanford University; Joshua B. Tenenbaum, a professor in MIT’s Department of Brain and Cognitive Sciences and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); Tomás Lozano-Pérez, an MIT professor of computer science and engineering and a member of CSAIL; and senior author Leslie Kaelbling, the Panasonic Professor of Computer Science and Engineering at MIT and a member of CSAIL. The research will be presented at the Conference on Robot Learning.

Constraint complications

Continuous constraint satisfaction problems are particularly challenging for robots. These problems appear in multistep robot manipulation tasks, like packing items into a box or setting a dinner table. They often involve achieving a number of constraints, including geometric constraints, such as avoiding collisions between the robot arm and the environment; physical constraints, such as stacking objects so they are stable; and qualitative constraints, such as placing a spoon to the right of a knife.

There may be many constraints, and they vary across problems and environments depending on the geometry of objects and human-specified requirements.

To solve these problems efficiently, the MIT researchers developed a machine-learning technique called Diffusion-CCSP. Diffusion models learn to generate new data samples that resemble samples in a training dataset by iteratively refining their output.

To do this, diffusion models learn a procedure for making small improvements to a potential solution. Then, to solve a problem, they start with a random, very bad solution and then gradually improve it.

Using generative AI models, MIT researchers created a technique that could enable robots to efficiently solve continuous constraint satisfaction problems, such as packing objects into a box while avoiding collisions, as shown in this simulation. Image: Courtesy of the researchers.

For example, imagine randomly placing plates and utensils on a simulated table, allowing them to physically overlap. The collision-free constraints between objects will result in them nudging each other away, while qualitative constraints will drag the plate to the center, align the salad fork and dinner fork, etc.

Diffusion models are well-suited for this kind of continuous constraint-satisfaction problem because the influences from multiple models on the pose of one object can be composed to encourage the satisfaction of all constraints, Yang explains. By starting from a random initial guess each time, the models can obtain a diverse set of good solutions.

Working together

For Diffusion-CCSP, the researchers wanted to capture the interconnectedness of the constraints. In packing for instance, one constraint might require a certain object to be next to another object, while a second constraint might specify where one of those objects must be located.

Diffusion-CCSP learns a family of diffusion models, with one for each type of constraint. The models are trained together, so they share some knowledge, like the geometry of the objects to be packed.

The models then work together to find solutions, in this case locations for the objects to be placed, that jointly satisfy the constraints.

“We don’t always get to a solution at the first guess. But when you keep refining the solution and some violation happens, it should lead you to a better solution. You get guidance from getting something wrong,” she says.

Training individual models for each constraint type and then combining them to make predictions greatly reduces the amount of training data required, compared to other approaches.

However, training these models still requires a large amount of data that demonstrate solved problems. Humans would need to solve each problem with traditional slow methods, making the cost to generate such data prohibitive, Yang says.

Instead, the researchers reversed the process by coming up with solutions first. They used fast algorithms to generate segmented boxes and fit a diverse set of 3D objects into each segment, ensuring tight packing, stable poses, and collision-free solutions.

“With this process, data generation is almost instantaneous in simulation. We can generate tens of thousands of environments where we know the problems are solvable,” she says.

Trained using these data, the diffusion models work together to determine locations objects should be placed by the robotic gripper that achieve the packing task while meeting all of the constraints.

They conducted feasibility studies, and then demonstrated Diffusion-CCSP with a real robot solving a number of difficult problems, including fitting 2D triangles into a box, packing 2D shapes with spatial relationship constraints, stacking 3D objects with stability constraints, and packing 3D objects with a robotic arm.

This figure shows examples of 2D triangle packing. These are collision-free configurations. Image: courtesy of the researchers.

This figure shows 3D object stacking with stability constraints. Researchers say at least one object is supported by multiple objects. Image: courtesy of the researchers.

Their method outperformed other techniques in many experiments, generating a greater number of effective solutions that were both stable and collision-free.

In the future, Yang and her collaborators want to test Diffusion-CCSP in more complicated situations, such as with robots that can move around a room. They also want to enable Diffusion-CCSP to tackle problems in different domains without the need to be retrained on new data.

“Diffusion-CCSP is a machine-learning solution that builds on existing powerful generative models,” says Danfei Xu, an assistant professor in the School of Interactive Computing at the Georgia Institute of Technology and a Research Scientist at NVIDIA AI, who was not involved with this work. “It can quickly generate solutions that simultaneously satisfy multiple constraints by composing known individual constraint models. Although it’s still in the early phases of development, the ongoing advancements in this approach hold the promise of enabling more efficient, safe, and reliable autonomous systems in various applications.”

This research was funded, in part, by the National Science Foundation, the Air Force Office of Scientific Research, the Office of Naval Research, the MIT-IBM Watson AI Lab, the MIT Quest for Intelligence, the Center for Brains, Minds, and Machines, Boston Dynamics Artificial Intelligence Institute, the Stanford Institute for Human-Centered Artificial Intelligence, Analog Devices, JPMorgan Chase and Co., and Salesforce.

• PAPER – Compositional Diffusion-Based Continuous Constraint Solvers. Zhutian Yang, Jiayuan Mao, Yilun Du, Jiajun Wu,
  Joshua B. Tenenbaum, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. arXive.

MIT researchers have developed a camera-based touch sensor that is long, curved, and shaped like a human finger. Their device, which provides high-resolution tactile sensing over a large area, could enable a robotic hand to perform multiple types of grasps. Image: Courtesy of the researchers

By Adam Zewe | MIT News

Imagine grasping a heavy object, like a pipe wrench, with one hand. You would likely grab the wrench using your entire fingers, not just your fingertips. Sensory receptors in your skin, which run along the entire length of each finger, would send information to your brain about the tool you are grasping.

In a robotic hand, tactile sensors that use cameras to obtain information about grasped objects are small and flat, so they are often located in the fingertips. These robots, in turn, use only their fingertips to grasp objects, typically with a pinching motion. This limits the manipulation tasks they can perform.

MIT researchers have developed a camera-based touch sensor that is long, curved, and shaped like a human finger. Their device provides high-resolution tactile sensing over a large area. The sensor, called the GelSight Svelte, uses two mirrors to reflect and refract light so that one camera, located in the base of the sensor, can see along the entire finger’s length.

In addition, the researchers built the finger-shaped sensor with a flexible backbone. By measuring how the backbone bends when the finger touches an object, they can estimate the force being placed on the sensor.

They used GelSight Svelte sensors to produce a robotic hand that was able to grasp a heavy object like a human would, using the entire sensing area of all three of its fingers. The hand could also perform the same pinch grasps common to traditional robotic grippers.

This gif shows a robotic hand that incorporates three, finger-shaped GelSight Svelte sensors. The sensors, which provide high-resolution tactile sensing over a large area, enable the hand to perform multiple grasps, including pinch grasps that use only the fingertips and a power grasp that uses the entire sensing area of all three fingers. Credit: Courtesy of the researchers

“Because our new sensor is human finger-shaped, we can use it to do different types of grasps for different tasks, instead of using pinch grasps for everything. There’s only so much you can do with a parallel jaw gripper. Our sensor really opens up some new possibilities on different manipulation tasks we could do with robots,” says Alan (Jialiang) Zhao, a mechanical engineering graduate student and lead author of a paper on GelSight Svelte.

Zhao wrote the paper with senior author Edward Adelson, the John and Dorothy Wilson Professor of Vision Science in the Department of Brain and Cognitive Sciences and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research will be presented at the IEEE Conference on Intelligent Robots and Systems.

Mirror mirror

Cameras used in tactile sensors are limited by their size, the focal distance of their lenses, and their viewing angles. Therefore, these tactile sensors tend to be small and flat, which confines them to a robot’s fingertips.

With a longer sensing area, one that more closely resembles a human finger, the camera would need to sit farther from the sensing surface to see the entire area. This is particularly challenging due to size and shape restrictions of a robotic gripper.

Zhao and Adelson solved this problem using two mirrors that reflect and refract light toward a single camera located at the base of the finger.

GelSight Svelte incorporates one flat, angled mirror that sits across from the camera and one long, curved mirror that sits along the back of the sensor. These mirrors redistribute light rays from the camera in such a way that the camera can see the along the entire finger’s length.

To optimize the shape, angle, and curvature of the mirrors, the researchers designed software to simulate reflection and refraction of light.

“With this software, we can easily play around with where the mirrors are located and how they are curved to get a sense of how well the image will look after we actually make the sensor,” Zhao explains.

The mirrors, camera, and two sets of LEDs for illumination are attached to a plastic backbone and encased in a flexible skin made from silicone gel. The camera views the back of the skin from the inside; based on the deformation, it can see where contact occurs and measure the geometry of the object’s contact surface.

A breakdown of the components that make up the finger-like touch sensor. Image: Courtesy of the researchers

In addition, the red and green LED arrays give a sense of how deeply the gel is being pressed down when an object is grasped, due to the saturation of color at different locations on the sensor.

The researchers can use this color saturation information to reconstruct a 3D depth image of the object being grasped.

The sensor’s plastic backbone enables it to determine proprioceptive information, such as the twisting torques applied to the finger. The backbone bends and flexes when an object is grasped. The researchers use machine learning to estimate how much force is being applied to the sensor, based on these backbone deformations.

However, combining these elements into a working sensor was no easy task, Zhao says.

“Making sure you have the correct curvature for the mirror to match what we have in simulation is pretty challenging. Plus, I realized there are some kinds of superglue that inhibit the curing of silicon. It took a lot of experiments to make a sensor that actually works,” he adds.

Versatile grasping

Once they had perfected the design, the researchers tested the GelSight Svelte by pressing objects, like a screw, to different locations on the sensor to check image clarity and see how well it could determine the shape of the object.

They also used three sensors to build a GelSight Svelte hand that can perform multiple grasps, including a pinch grasp, lateral pinch grasp, and a power grasp that uses the entire sensing area of the three fingers. Most robotic hands, which are shaped like parallel jaw drippers, can only perform pinch grasps.

A three-finger power grasp enables a robotic hand to hold a heavier object more stably. However, pinch grasps are still useful when an object is very small. Being able to perform both types of grasps with one hand would give a robot more versatility, he says.

Moving forward, the researchers plan to enhance the GelSight Svelte so the sensor is articulated and can bend at the joints, more like a human finger.

“Optical-tactile finger sensors allow robots to use inexpensive cameras to collect high-resolution images of surface contact, and by observing the deformation of a flexible surface the robot estimates the contact shape and forces applied. This work represents an advancement on the GelSight finger design, with improvements in full-finger coverage and the ability to approximate bending deflection torques using image differences and machine learning,” says Monroe Kennedy III, assistant professor of mechanical engineering at Stanford University, who was not involved with this research. “Improving a robot’s sense of touch to approach human ability is a necessity and perhaps the catalyst problem for developing robots capable of working on complex, dexterous tasks.”

This research is supported, in part, by the Toyota Research Institute.

• PAPER – GelSight Svelte: A Human Finger-shaped Single-camera Tactile Robot Finger with Large Sensing Coverage and Proprioceptive Sensing. Jialiang Zhao and Edward H. Adelson. arXive.

• PAPER – GelSight Svelte Hand: A Three-finger, Two-DoF, Tactile-rich, Low-cost Robot Hand for Dexterous Manipulation. Jialiang Zhao and Edward H. Adelson. arXive.

Scientists at the Max Planck Institute for Intelligent Systems in Stuttgart have developed a soft robotic tool that promises to one day transform minimally invasive endovascular surgery. The two-part magnetic tool can help to visualise in real time the fine morphological details of partial vascular blockages such as stenoses, even in the narrowest and most curved vessels. It can also find its way through severe blockages such as chronic total occlusions. This tool could one day take the perception of endovascular medical devices a step further.

Intravascular imaging techniques and microcatheter procedures are becoming ever more advanced, revolutionizing the diagnosis and treatment of many diseases. However, current methods often fail to accurately detect the fine features of vascular disease, such as those seen from within occluded vessels, due to limitations such as uneven contrast agent diffusion and difficulty in safely accessing occluded vessels. Such limitations can delay rapid intervention and treatment of a patient.

Scientists at the Max Planck Institute for Intelligent Systems in Stuttgart have looked at this problem. They have leveraged the concepts of soft robotics and microfabrication to develop a miniature soft magnetic tool that looks like a very slim eel. This tool may one day take the perception capabilities of endovascular devices one step further. In a paper and in a video, the team shows how the tool, which is propelled forward by the blood flow, travels through the narrowest artificial vessels – whether there is a sharp bend, curve, or obstacle.

When the tool reaches an occlusion like a partially blocked artery, it performs a wave-like deformation given the external magnetic field (more on that below). Then, the deformed soft body will be gently in contact with the surrounding occluded structures. Lastly, the real-time shapes of the device when we retract it will ‘visualize’ the morphological details inside the vessel, which facilitates the drug release at occlusion, as well as the sizing and placement of medical devices like stents and balloons for following treatment.

When there is a severe occlusion with only tiny microchannels for the blood to flow through, the tool can utilize the force from the blood to easily slide through these narrow channels. Which way was chosen indicates to the surgeon which access route to take for the following medical operation.

“The methods of diagnosing and treating endovascular narrow diseases such as vascular stenosis or chronic total occlusion are still very limited. It is difficult to accurately detect and cross these areas in the very complex network of vessels inside the body”, says Yingbo Yan, who is a guest researcher in the Physical Intelligence Department at MPI-IS. He is the first author of the paper “Magnetically-assisted soft milli-tools for occluded lumen morphology detection”, which was published in Science Advances on August 18, 2023. “We hope that our new soft robotic tool can one day help accurately detect and navigate through the many complex and narrow vessels inside a body, and perform treatments more effectively, reducing potential risks.”

This tiny and soft tool has a 20 mm long magnetic Active Deformation Segment (ADS) and a 5mm long Fluid Drag-driven Segment (FDS). The magnetization profile of ADS is pre-programmed with a vibrating-sample magnetometer, providing a uniform magnetic field. Under an external magnetic field, this part can deform into a sinusoidal shape, easily adapting to the surrounding environment and deforming into various shapes. Thus, continuous monitoring of the shape changes of ADS while retracting it can provide detailed morphological information of the partial occlusions inside a vessel.

The FDS was fabricated using a soft polymer. Small beams on its side are bent by the fluidic drag from the incoming flow. In this way, the entire tool is carried towards the area with the highest flow velocity. Therefore, learning the location of the FDS while advancing it can point to the location and the route of the microchannel inside the severe occlusions.

“Detection of vascular diseases in the distal and hard-to-reach vascular regions such as the brain can be more challenging clinically, and our tool could work with Stentbot in the untethered mode”, says Tianlu Wang, a postdoc in the Physical Intelligence Department at MPI-IS and another first author of the work. “Stentbot is a wireless robot used for locomotion and medical functions in the distal vasculature we recently developed in our research group. We believe this new soft robotic tool can add new capabilities to wireless robots and contribute new solutions in these challenging regions.”

“Our tool shows potential to greatly improve minimally invasive medicine. This technology can reach and detect areas that were previously difficult to access. We expect that our robot can help make the diagnosis and treatment of, for instance, stenosis or a CTO more precise and timelier”, says Metin Sitti, Director of the Physical Intelligence Department at MPI-IS, Professor at Koç University and ETH Zurich.

• PAPER – Magnetically assisted soft milli-tools for occluded lumen morphology detection. Yingbo Yan, Tianlu Wang, Rongjing Zhang, Yilun Liu, Wenqi Hu and Metin Sitti. Science Advances, 9(33), eadi3979.

By Matt Shipman

Researchers who created a soft robot that could navigate simple mazes without human or computer direction have now built on that work, creating a “brainless” soft robot that can navigate more complex and dynamic environments.

“In our earlier work, we demonstrated that our soft robot was able to twist and turn its way through a very simple obstacle course,” says Jie Yin, co-corresponding author of a paper on the work and an associate professor of mechanical and aerospace engineering at North Carolina State University. “However, it was unable to turn unless it encountered an obstacle. In practical terms this meant that the robot could sometimes get stuck, bouncing back and forth between parallel obstacles.

“We’ve developed a new soft robot that is capable of turning on its own, allowing it to make its way through twisty mazes, even negotiating its way around moving obstacles. And it’s all done using physical intelligence, rather than being guided by a computer.”

Physical intelligence refers to dynamic objects – like soft robots – whose behavior is governed by their structural design and the materials they are made of, rather than being directed by a computer or human intervention.

As with the earlier version, the new soft robots are made of ribbon-like liquid crystal elastomers. When the robots are placed on a surface that is at least 55 degrees Celsius (131 degrees Fahrenheit), which is hotter than the ambient air, the portion of the ribbon touching the surface contracts, while the portion of the ribbon exposed to the air does not. This induces a rolling motion; the warmer the surface, the faster the robot rolls.

However, while the previous version of the soft robot had a symmetrical design, the new robot has two distinct halves. One half of the robot is shaped like a twisted ribbon that extends in a straight line, while the other half is shaped like a more tightly twisted ribbon that also twists around itself like a spiral staircase.

This asymmetrical design means that one end of the robot exerts more force on the ground than the other end. Think of a plastic cup that has a mouth wider than its base. If you roll it across the table, it doesn’t roll in a straight line – it makes an arc as it travels across the table. That’s due to its asymmetrical shape.

“The concept behind our new robot is fairly simple: because of its asymmetrical design, it turns without having to come into contact with an object,” says Yao Zhao, first author of the paper and a postdoctoral researcher at NC State. “So, while it still changes directions when it does come into contact with an object – allowing it to navigate mazes – it cannot get stuck between parallel objects. Instead, its ability to move in arcs allows it to essentially wiggle its way free.”

The researchers demonstrated the ability of the asymmetrical soft robot design to navigate more complex mazes – including mazes with moving walls – and fit through spaces narrower than its body size. The researchers tested the new robot design on both a metal surface and in sand.

“This work is another step forward in helping us develop innovative approaches to soft robot design – particularly for applications where soft robots would be able to harvest heat energy from their environment,” Yin says.

The paper, “Physically Intelligent Autonomous Soft Robotic Maze Escaper,” appears in the journal Science Advances. First author of the paper is Yao Zhao, a postdoctoral researcher at NC State. Hao Su, an associate professor of mechanical and aerospace engineering at NC State, is co-corresponding author. Additional co-authors include Yaoye Hong, a recent Ph.D. graduate of NC State; Yanbin Li, a postdoctoral researcher at NC State; and Fangjie Qi and Haitao Qing, both Ph.D. students at NC State.

The work was done with support from the National Science Foundation under grants 2005374, 2126072, 1944655 and 2026622.

• PAPER – Physically Intelligent Autonomous Soft Robotic Maze Escaper. Yao Zhao, Yaoye Hong, Yanbin Li, Fangjie Qi, Haitao Qing and Jie Yin. Science Advances, 9(36), eadi3254.

Researchers at the University of Washington developed small robotic devices that can change how they move through the air by “snapping” into a folded position during their descent. Shown here is a timelapse photo of the “microflier” falling in its unfolded state, which makes it tumble chaotically and spread outward in the wind. Photo by Mark Stone/University of Washington

By Roger Van Scyoc

On a cool afternoon at the heart of the University of Washington’s campus, autumn, for a few fleeting moments, appears to have arrived early. Tiny golden squares resembling leaves flutter then fall, switching from a frenzied tumble to a graceful descent with a snap.

Aptly named “microfliers” and inspired by Miura-fold origami, these small robotic devices can fold closed during their descent after being dropped from a drone. This “snapping” action changes the way they disperse and may, in the future, help change the way scientists study agriculture, meteorology, climate change and more.

“In nature, you see leaves and seeds disperse in just one manner,” said Kyle Johnson, an Allen School Ph.D. student and a first co-author of the paper on the subject published in Science Robotics. “What we were able to achieve was a structure that can actually act in two different ways.”

When open flat, the devices tumble chaotically, mimicking the descent of an elm leaf. When folded closed, they drop in a more stable manner, mirroring how a maple leaf falls from a branch. Through a number of methods — onboard pressure sensor, timer or a Bluetooth signal — the researchers can control when the devices transition from open to closed, and in doing so, manipulate how far they disperse through the air.

How could they achieve this? By reading between the lines.

“The Miura-ori origami fold, inspired by geometric patterns found in leaves, enables the creation of structures that can ‘snap’ between a flat and more folded state,” said co-senior author Vikram Iyer, an Allen School professor and co-director of the Computing for the Environment (CS4Env) initiative. “Because it only takes energy to switch between the states, we began exploring this as an energy efficient way to change surface area in mid-air, with the intuition that opening or closing a parachute will change how fast an object falls.”

That energy efficiency is key to being able to operate without batteries and scale down the fliers’ size and weight. Fitted with a battery-free actuator and a solar power-harvesting circuit, microfliers boast energy-saving features not seen in larger and heavier battery-powered counterparts such as drones. Yet they are robust enough to carry sensors for a number of metrics, including temperature, pressure, humidity and altitude. Beyond measuring atmospheric conditions, the researchers say a network of these devices could help paint a picture of crop growth on farmland or detect gas leaks near population centers.

“This approach opens up a new design space for microfliers by using origami,” said Shyam Gollakota, the Thomas J. Cable Endowed Professor in the Allen School and director of the school’s Mobile Intelligence Lab who was also a co-senior author. “We hope this work is the first step towards a future vision for creating a new class of fliers and flight modalities.”

Weighing less than half a gram, microfliers require less material and cost less than drones. They also offer the ability to go where it’s too dangerous for a human to set foot.

For instance, Johnson said, microfliers could be deployed when tracking forest fires. Currently, firefighting teams sometimes rappel down to where a fire is spreading. Microfliers could assist in mapping where a fire may be heading and where best to drop a payload of water. Furthermore, the team is working on making more components of the device biodegradable in the case that they can’t be recovered after being released.

“There’s a good amount of work toward making these circuits more sustainable,” said Vicente Arroyos, another Allen School Ph.D. student and first co-author on the paper. “We can leverage our work on biodegradable materials to make these more sustainable.”

Besides improving sustainability, the researchers also tackled challenges relating to the structure of the device itself. Early prototypes lacked the carbon fiber roots that provide the rigidity needed to prevent accidental transitions between states.

The research team took inspiration from elm and maple leaves in designing the microfliers. When open flat, the devices tumble chaotically, similar to how an elm leaf falls from a branch. When they are “snapped” into a folded position, as shown here, they descend in a more stable, straight downward manner like a maple leaf. Photo by Mark Stone/University of Washington

Collecting maple and elm leaves from outside their lab, the researchers noticed that while their origami structures exhibited the bistability required to change between states, they flexed too easily and didn’t have the venation seen in the found foliage. To gain more fine-grained control, they took another cue from the environment.

“We looked again to nature to make the faces of the origami flat and rigid, adding a vein-like pattern to the structure using carbon fiber,” Johnson said. “After that modification, we no longer saw a lot of the energy that we input dissipate over the origami’s faces.”

In total, the researchers estimate that the development of their design took about two years. There’s still room to grow, they added, noting that the current microfliers can only transition from open to closed. They said newer designs, by offering the ability to switch back and forth between states, may offer more precision and flexibility in where and how they’re used.

During testing, when dropped from an altitude of 40 meters, for instance, the microfliers could disperse up to distances of 98 meters in a light breeze. Further refinements could increase the area of coverage, allowing them to follow more precise trajectories by accounting for variables such as wind and inclement conditions.

Related to their previous work with dandelion-inspired sensors, the origami microfliers build upon the researchers’ larger goal of creating the internet of bio-inspired things. Whereas the dandelion-inspired devices featured passive flight, reflecting the manner in which dandelion seeds disperse through the wind, the origami microfliers function as complete robotic systems that include actuation to change their shape, active and bi-directional wireless transmission via an onboard radio, and onboard computing and sensing to autonomously trigger shape changes upon reaching a target altitude.

“This design can also accommodate additional sensors and payload due to its size and power harvesting capabilities,” Arroyos said. “It’s exciting to think about the untapped potential for these devices.”

The future, in other words, is quickly taking shape.

“Origami is inspired by nature,” Johnson added, smiling. “These patterns are all around us. We just have to look in the right place.”

The project was an interdisciplinary work by an all-UW team. The paper’s co-authors also included Amélie Ferran, a Ph.D. student in the mechanical engineering department, as well as Raul Villanueva, Dennis Yin and Tilboon Elberier, who contributed as undergraduate students studying electrical and computer engineering, and mechanical engineering professors Alberto Aliseda and Sawyer Fuller.

Johnson and Arroyos, who co-founded and currently lead the educational nonprofit AVELA – A Vision for Engineering Literacy & Access, and their teammates have done outreach efforts in Washington state K-12 schools related to the research, including showing students how to create their own bi-stable leaf-out origami structure using a piece of paper. Check out a related demonstration video here, and learn more about the microflier project here and in a related UW News release and GeekWire story.

• PAPER – Solar-powered shape-changing origami microfliers. Kyle Johnson, Vicente Arroyos, Amélie Ferran, Raul Villanueva, Dennis Yin, Alberto Aliseda, Sawyer Fuller, Vikram Iyer, and Shyamnath Gollakota. Science Robotics 8.82 (2023): eadg4276.

By Deborah Pirchner

Malaria is an infectious disease claiming more than half a million lives each year. Because traditional diagnosis takes expertise and the workload is high, an international team of researchers investigated if diagnosis using a new system combining an automatic scanning microscope and AI is feasible in clinical settings. They found that the system identified malaria parasites almost as accurately as experts staffing microscopes used in standard diagnostic procedures. This may help reduce the burden on microscopists and increase the feasible patient load.

Each year, more than 200 million people fall sick with malaria and more than half a million of these infections lead to death. The World Health Organization recommends parasite-based diagnosis before starting treatment for the disease caused by Plasmodium parasites. There are various diagnostic methods, including conventional light microscopy, rapid diagnostic tests and PCR.

The standard for malaria diagnosis, however, remains manual light microscopy, during which a specialist examines blood films with a microscope to confirm the presence of malaria parasites. Yet, the accuracy of the results depends critically on the skills of the microscopist and can be hampered by fatigue caused by excessive workloads of the professionals doing the testing.

Now, writing in Frontiers in Malaria, an international team of researchers has assessed whether a fully automated system, combining AI detection software and an automated microscope, can diagnose malaria with clinically useful accuracy.

“At an 88% diagnostic accuracy rate relative to microscopists, the AI system identified malaria parasites almost, though not quite, as well as experts,” said Dr Roxanne Rees-Channer, a researcher at The Hospital for Tropical Diseases at UCLH in the UK, where the study was performed. “This level of performance in a clinical setting is a major achievement for AI algorithms targeting malaria. It indicates that the system can indeed be a clinically useful tool for malaria diagnosis in appropriate settings.”

AI delivers accurate diagnosis

The researchers sampled more than 1,200 blood samples of travelers who had returned to the UK from malaria-endemic countries. The study tested the accuracy of the AI and automated microscope system in a true clinical setting under ideal conditions.

They evaluated samples using both manual light microscopy and the AI-microscope system. By hand, 113 samples were diagnosed as malaria parasite positive, whereas the AI-system correctly identified 99 samples as positive, which corresponds to an 88% accuracy rate.

“AI for medicine often posts rosy preliminary results on internal datasets, but then falls flat in real clinical settings. This study independently assessed whether the AI system could succeed in a true clinical use case,” said Rees-Channer, who is also the lead author of the study.

Automated vs manual

The fully automated malaria diagnostic system the researchers put to the test includes hard- as well as software. An automated microscopy platform scans blood films and malaria detection algorithms process the image to detect parasites and the quantity present.

Automated malaria diagnosis has several potential benefits, the scientists pointed out. “Even expert microscopists can become fatigued and make mistakes, especially under a heavy workload,” Rees-Channer explained. “Automated diagnosis of malaria using AI could reduce this burden for microscopists and thus increase the feasible patient load.” Furthermore, these systems deliver reproducible results and can be widely deployed, the scientists wrote.

Despite the 88% accuracy rate, the automated system also falsely identified 122 samples as positive, which can lead to patients receiving unnecessary anti-malarial drugs. “The AI software is still not as accurate as an expert microscopist. This study represents a promising datapoint rather than a decisive proof of fitness,” Rees-Channer concluded.

Read the research in full

Evaluation of an automated microscope using machine learning for the detection of malaria in travelers returned to the UK, Roxanne R. Rees-Channer, Christine M. Bachman, Lynn Grignard, Michelle L. Gatton, Stephen Burkot, Matthew P. Horning, Charles B. Delahunt, Liming Hu, Courosh Mehanian, Clay M. Thompson, Katherine Woods, Paul Lansdell, Sonal Shah, Peter L. Chiodini, Frontiers in Malaria (2023).

This illustration shows a 3D printed heart ventricle engineered with fiber-infused ink. Credit: Harvard SEAS

By Kat J. McAlpine / SEAS Communications

Over the last decade, advances in 3D printing have unlocked new possibilities for bioengineers to build heart tissues and structures. Their goals include creating better in vitro platforms for discovering new therapeutics for heart disease, the leading cause of death in the United States, responsible for about one in every five deaths nationally, and using 3D-printed cardiac tissues to evaluate which treatments might work best in individual patients. A more distant aim is to fabricate implantable tissues that can heal or replace faulty or diseased structures inside a patient’s heart.

In a paper published in Nature Materials, researchers from Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) and the Wyss Institute for Biologically Inspired Engineering at Harvard University report the development of a new hydrogel ink infused with gelatin fibers that enables 3D printing of a functional heart ventricle that mimics beating like a human heart. They discovered the fiber-infused gel (FIG) ink allows heart muscle cells printed in the shape of a ventricle to align and beat in coordination like a human heart chamber.

“People have been trying to replicate organ structures and functions to test drug safety and efficacy as a way of predicting what might happen in the clinical setting,” says Suji Choi, research associate at SEAS and first author on the paper. But until now, 3D printing techniques alone have not been able to achieve physiologically-relevant alignment of cardiomyocytes, the cells responsible for transmitting electrical signals in a coordinated fashion to contract heart muscle.

“We started this project to address some of the inadequacies in 3D printing of biological tissues.”

– Kevin “Kit” Parker

The innovation lies in the addition of fibers within a printable ink. “FIG ink is capable of flowing through the printing nozzle but, once the structure is printed, it maintains its 3D shape,” says Choi. “Because of those properties, I found it’s possible to print a ventricle-like structure and other complex 3D shapes without using extra support materials or scaffolds.”

This video shows the spontaneous beating of a 3D-printed heart muscle. Credit: Harvard SEAS.

To create the FIG ink, Choi leveraged a rotary jet spinning technique developed in the lab of Kevin “Kit” Parker, Ph.D. that fabricates microfiber materials using an approach similar to the way cotton candy is spun. Postdoctoral researcher and Wyss Lumineer Luke MacQueen, a co-author on the paper, proposed the idea that fibers created by the rotary jet spinning technique could be added to an ink and 3D printed. Parker is a Wyss Associate Faculty member and the Tarr Family Professor of Bioengineering and Applied Physics at SEAS.

“When Luke developed this concept, the vision was to broaden the range of spatial scales that could be printed with 3D printers by dropping the bottom out of the lower limits, taking it down to the nanometer scale,” Parker says. “The advantage of producing the fibers with rotary jet spinning rather than electrospinning” – a more conventional method for generating ultrathin fibers – “is that we can use proteins that would otherwise be degraded by the electrical fields in electrospinning.”

Using the rotary jet to spin gelatin fibers, Choi produced a sheet of material with a similar appearance to cotton. Next, she used sonification – sound waves – to break that sheet into fibers about 80 to 100 micrometers long and about 5 to 10 micrometers in diameter. Then, she dispersed those fibers into a hydrogel ink.

“This concept is broadly applicable – we can use our fiber-spinning technique to reliably produce fibers in the lengths and shapes we want.”

– Suji Choi

The most difficult aspect was troubleshooting the desired ratio between fibers and hydrogel in the ink to maintain fiber alignment and the overall integrity of the 3D-printed structure.

As Choi printed 2D and 3D structures using FIG ink, the cardiomyocytes lined up in tandem with the direction of the fibers inside the ink. By controlling the printing direction, Choi could therefore control how the heart muscle cells would align.

The tissue-engineered 3D ventricle model. Credit: Harvard SEAS

When she applied electrical stimulation to 3D-printed structures made with FIG ink, she found it triggered a coordinated wave of contractions in alignment with the direction of those fibers. In a ventricle-shaped structure, “it was very exciting to see the chamber actually pumping in a similar way to how real heart ventricles pump,” Choi says.

As she experimented with more printing directions and ink formulas, she found she could generate even stronger contractions within ventricle-like shapes.

“Compared to the real heart, our ventricle model is simplified and miniaturized,” she says. The team is now working toward building more life-like heart tissues with thicker muscle walls that can pump fluid more strongly. Despite not being as strong as real heart tissue, the 3D-printed ventricle could pump 5-20 times more fluid volume than previous 3D-printed heart chambers.

The team says the technique can also be used to build heart valves, dual-chambered miniature hearts, and more.

“FIGs are but one tool we have developed for additive manufacturing,” Parker says. “We have other methods in development as we continue our quest to build human tissues for regenerative therapeutics. The goal is not to be tool driven – we are tool agnostic in our search for a better way to build biology.”

Additional authors include Keel Yong Lee, Sean L. Kim, Huibin Chang, John F. Zimmerman, Qianru Jin, Michael M. Peters, Herdeline Ann M. Ardoña, Xujie Liu, Ann-Caroline Heiler, Rudy Gabardi, Collin Richardson, William T. Pu, and Andreas Bausch.

This work was sponsored by SEAS; the National Science Foundation through the Harvard University Materials Research Science and Engineering Center (DMR-1420570, DMR-2011754); the National Institutes of Health and National Center for Advancing Translational Sciences (UH3HL141798, 225 UG3TR003279); the Harvard University Center for Nanoscale Systems (CNS), a member of the National Nanotechnology Coordinated Infrastructure Network (NNCI) which is supported by the National Science Foundation (ECCS-2025158, S10OD023519); and the American Chemical Society’s Irving S. Sigal Postdoctoral Fellowships.

• PAPER – Fibre-infused gel scaffolds guide cardiomyocyte alignment in 3D-printed ventricles. Suji Choi, Keel Yong Lee, Sean L. Kim, Luke A. MacQueen, Huibin Chang, John F. Zimmerman, Qianru Jin, Michael M. Peters, Herdeline Ann M. Ardoña, Xujie Liu, Ann-Caroline Heiler, Rudy Gabardi, Collin Richardson, William T. Pu, Andreas R. Bausch and Kevin Kit Parker. Nat. Mater. 22, 1039–1046 (2023). https://doi.org/10.1038/s41563-023-01611-3

MIT researchers used kirigami, the art of Japanese paper cutting and folding, to develop ultrastrong, lightweight materials that have tunable mechanical properties, like stiffness and flexibility. These materials could be used in airplanes, automobiles, or spacecraft. Image: Courtesy of the researchers

By Adam Zewe | MIT News

Cellular solids are materials composed of many cells that have been packed together, such as a honeycomb. The shape of those cells largely determines the material’s mechanical properties, including its stiffness or strength. Bones, for instance, are filled with a natural material that enables them to be lightweight, but stiff and strong.

Inspired by bones and other cellular solids found in nature, humans have used the same concept to develop architected materials. By changing the geometry of the unit cells that make up these materials, researchers can customize the material’s mechanical, thermal, or acoustic properties. Architected materials are used in many applications, from shock-absorbing packing foam to heat-regulating radiators.

Using kirigami, the ancient Japanese art of folding and cutting paper, MIT researchers have now manufactured a type of high-performance architected material known as a plate lattice, on a much larger scale than scientists have previously been able to achieve by additive fabrication. This technique allows them to create these structures from metal or other materials with custom shapes and specifically tailored mechanical properties. 

“This material is like steel cork. It is lighter than cork, but with high strength and high stiffness,” says Professor Neil Gershenfeld, who leads the Center for Bits and Atoms (CBA) at MIT and is senior author of a new paper on this approach.

The researchers developed a modular construction process in which many smaller components are formed, folded, and assembled into 3D shapes. Using this method, they fabricated ultralight and ultrastrong structures and robots that, under a specified load, can morph and hold their shape.

Because these structures are lightweight but strong, stiff, and relatively easy to mass-produce at larger scales, they could be especially useful in architectural, airplane, automotive, or aerospace components.

Joining Gershenfeld on the paper are co-lead authors Alfonso Parra Rubio, a research assistant in the CBA, and Klara Mundilova, an MIT electrical engineering and computer science graduate student; along with David Preiss, a graduate student in the CBA; and Erik D. Demaine, an MIT professor of computer science. The research will be presented at ASME’s Computers and Information in Engineering Conference.

The researchers actuate a corrugated structure by tensioning steel wires across the compliant surfaces and then connecting them to a system of pulleys and motors, enabling the structure to bend in either direction. Image: Courtesy of the researchers

Fabricating by folding

Architected materials, like lattices, are often used as cores for a type of composite material known as a sandwich structure. To envision a sandwich structure, think of an airplane wing, where a series of intersecting, diagonal beams form a lattice core that is sandwiched between a top and bottom panel. This truss lattice has high stiffness and strength, yet is very lightweight.

Plate lattices are cellular structures made from three-dimensional intersections of plates, rather than beams. These high-performance structures are even stronger and stiffer than truss lattices, but their complex shape makes them challenging to fabricate using common techniques like 3D printing, especially for large-scale engineering applications.

The MIT researchers overcame these manufacturing challenges using kirigami, a technique for making 3D shapes by folding and cutting paper that traces its history to Japanese artists in the 7th century.

Kirigami has been used to produce plate lattices from partially folded zigzag creases. But to make a sandwich structure, one must attach flat plates to the top and bottom of this corrugated core onto the narrow points formed by the zigzag creases. This often requires strong adhesives or welding techniques that can make assembly slow, costly, and challenging to scale.

The MIT researchers modified a common origami crease pattern, known as a Miura-ori pattern, so the sharp points of the corrugated structure are transformed into facets. The facets, like those on a diamond, provide flat surfaces to which the plates can be attached more easily, with bolts or rivets.

The MIT researchers modified a common origami crease pattern, known as a Miura-ori pattern, so the sharp points of the corrugated structure are transformed into facets. The facets, like those on a diamond, provide flat surfaces to which the plates can be attached more easily, with bolts or rivets. Image: Courtesy of the researchers

“Plate lattices outperform beam lattices in strength and stiffness while maintaining the same weight and internal structure,” says Parra Rubio. “Reaching the H-S upper bound for theoretical stiffness and strength has been demonstrated through nanoscale production using two-photon lithography. Plate lattices construction has been so difficult that there has been little research on the macro scale. We think folding is a path to easier utilization of this type of plate structure made from metals.”

Customizable properties

Moreover, the way the researchers design, fold, and cut the pattern enables them to tune certain mechanical properties, such as stiffness, strength, and flexural modulus (the tendency of a material to resist bending). They encode this information, as well as the 3D shape, into a creasing map that is used to create these kirigami corrugations.

For instance, based on the way the folds are designed, some cells can be shaped so they hold their shape when compressed while others can be modified so they bend. In this way, the researchers can precisely control how different areas of the structure will deform when compressed.

Because the flexibility of the structure can be controlled, these corrugations could be used in robots or other dynamic applications with parts that move, twist, and bend.

To craft larger structures like robots, the researchers introduced a modular assembly process. They mass produce smaller crease patterns and assemble them into ultralight and ultrastrong 3D structures. Smaller structures have fewer creases, which simplifies the manufacturing process.

Using the adapted Miura-ori pattern, the researchers create a crease pattern that will yield their desired shape and structural properties. Then they utilize a unique machine — a Zund cutting table — to score a flat, metal panel that they fold into the 3D shape.

“To make things like cars and airplanes, a huge investment goes into tooling. This manufacturing process is without tooling, like 3D printing. But unlike 3D printing, our process can set the limit for record material properties,” Gershenfeld says.

Using their method, they produced aluminum structures with a compression strength of more than 62 kilonewtons, but a weight of only 90 kilograms per square meter. (Cork weighs about 100 kilograms per square meter.) Their structures were so strong they could withstand three times as much force as a typical aluminum corrugation.

Using their method, researchers produced aluminum structures with a compression strength of more than 62 kilonewtons, but a weight of only 90 kilograms per square meter. Image: Courtesy of the researchers

The versatile technique could be used for many materials, such as steel and composites, making it well-suited for the production lightweight, shock-absorbing components for airplanes, automobiles, or spacecraft.

However, the researchers found that their method can be difficult to model. So, in the future, they plan to develop user-friendly CAD design tools for these kirigami plate lattice structures. In addition, they want to explore methods to reduce the computational costs of simulating a design that yields desired properties. 

“Kirigami corrugations holds exciting potential for architectural construction,” says James Coleman MArch ’14, SM ’14, co-founder of the design for fabrication and installation firm SumPoint, and former vice president for innovation and R&D at Zahner, who was not involved with this work. “In my experience producing complex architectural projects, current methods for constructing large-scale curved and doubly curved elements are material intensive and wasteful, and thus deemed impractical for most projects. While the authors’ technology offers novel solutions to the aerospace and automotive industries, I believe their cell-based method can also significantly impact the built environment. The ability to fabricate various plate lattice geometries with specific properties could enable higher performing and more expressive buildings with less material. Goodbye heavy steel and concrete structures, hello lightweight lattices!”

Parra Rubio, Mundilova and other MIT graduate students also used this technique to create three large-scale, folded artworks from aluminum composite that are on display at the MIT Media Lab. Despite the fact that each artwork is several meters in length, the structures only took a few hours to fabricate.

“At the end of the day, the artistic piece is only possible because of the math and engineering contributions we are showing in our papers. But we don’t want to ignore the aesthetic power of our work,” Parra Rubio says.

This work was funded, in part, by the Center for Bits and Atoms Research Consortia, an AAUW International Fellowship, and a GWI Fay Weber Grant.

• PAPER – Kirigami corrugations: strong, modular, and programmable plate lattice. Alfonso Parra Rubio, Klara Mundilova, David Preiss, Erik D. Demaine, and Neil Gershenfeld.

Dual arm robot holding crisp. Image: Yijiong Lin

The new Bi-Touch system, designed by scientists at the University of Bristol and based at the Bristol Robotics Laboratory, allows robots to carry out manual tasks by sensing what to do from a digital helper.

The findings, published in IEEE Robotics and Automation Letters, show how an AI agent interprets its environment through tactile and proprioceptive feedback, and then control the robots’ behaviours, enabling precise sensing, gentle interaction, and effective object manipulation to accomplish robotic tasks.

This development could revolutionise industries such as fruit picking, domestic service, and eventually recreate touch in artificial limbs.

Lead author Yijiong Lin from the Faculty of Engineering, explained: “With our Bi-Touch system, we can easily train AI agents in a virtual world within a couple of hours to achieve bimanual tasks that are tailored towards the touch. And more importantly, we can directly apply these agents from the virtual world to the real world without further training.

“The tactile bimanual agent can solve tasks even under unexpected perturbations and manipulate delicate objects in a gentle way.”

Bimanual manipulation with tactile feedback will be key to human-level robot dexterity. However, this topic is less explored than single-arm settings, partly due to the availability of suitable hardware along with the complexity of designing effective controllers for tasks with relatively large state-action spaces. The team were able to develop a tactile dual-arm robotic system using recent advances in AI and robotic tactile sensing.

The researchers built up a virtual world (simulation) that contained two robot arms equipped with tactile sensors. They then design reward functions and a goal-update mechanism that could encourage the robot agents to learn to achieve the bimanual tasks and developed a real-world tactile dual-arm robot system to which they could directly apply the agent.

The robot learns bimanual skills through Deep Reinforcement Learning (Deep-RL), one of the most advanced techniques in the field of robot learning. It is designed to teach robots to do things by letting them learn from trial and error akin to training a dog with rewards and punishments.

For robotic manipulation, the robot learns to make decisions by attempting various behaviours to achieve designated tasks, for example, lifting up objects without dropping or breaking them. When it succeeds, it gets a reward, and when it fails, it learns what not to do. With time, it figures out the best ways to grab things using these rewards and punishments. The AI agent is visually blind relying only on proprioceptive feedback – a body’s ability to sense movement, action and location and tactile feedback.

They were able to successfully enable to the dual arm robot to successfully safely lift items as fragile as a single Pringle crisp.

Co-author Professor Nathan Lepora added: “Our Bi-Touch system showcases a promising approach with affordable software and hardware for learning bimanual behaviours with touch in simulation, which can be directly applied to the real world. Our developed tactile dual-arm robot simulation allows further research on more different tasks as the code will be open-source, which is ideal for developing other downstream tasks.”

Yijiong concluded: “Our Bi-Touch system allows a tactile dual-arm robot to learn sorely from simulation, and to achieve various manipulation tasks in a gentle way in the real world.

“And now we can easily train AI agents in a virtual world within a couple of hours to achieve bimanual tasks that are tailored towards the touch.”

• PAPER – Bi-Touch: Bimanual Tactile Manipulation With Sim-to-Real Deep Reinforcement Learning. Yijiong Lin, Alex Church, Max Yang, Haoran Li, John Lloyd, Dandan Zhang, and Nathan F. Lepora. IEEE Robotics and Automation Letters, vol. 8, no. 9, pp. 5472-5479, Sept. 2023, doi: 10.1109/LRA.2023.3295991.

MIT researchers developed an AI technique that enables a robot to develop complex plans for manipulating an object using its entire hand, not just the fingertips. This model can generate effective plans in about a minute using a standard laptop. Here, a robot attempts to rotate a bucket 180 degrees. Image: Courtesy of the researchers

By Adam Zewe | MIT News

Imagine you want to carry a large, heavy box up a flight of stairs. You might spread your fingers out and lift that box with both hands, then hold it on top of your forearms and balance it against your chest, using your whole body to manipulate the box. 

Humans are generally good at whole-body manipulation, but robots struggle with such tasks. To the robot, each spot where the box could touch any point on the carrier’s fingers, arms, and torso represents a contact event that it must reason about. With billions of potential contact events, planning for this task quickly becomes intractable.

Now MIT researchers found a way to simplify this process, known as contact-rich manipulation planning. They use an AI technique called smoothing, which summarizes many contact events into a smaller number of decisions, to enable even a simple algorithm to quickly identify an effective manipulation plan for the robot.

While still in its early days, this method could potentially enable factories to use smaller, mobile robots that can manipulate objects with their entire arms or bodies, rather than large robotic arms that can only grasp using fingertips. This may help reduce energy consumption and drive down costs. In addition, this technique could be useful in robots sent on exploration missions to Mars or other solar system bodies, since they could adapt to the environment quickly using only an onboard computer.      

“Rather than thinking about this as a black-box system, if we can leverage the structure of these kinds of robotic systems using models, there is an opportunity to accelerate the whole procedure of trying to make these decisions and come up with contact-rich plans,” says H.J. Terry Suh, an electrical engineering and computer science (EECS) graduate student and co-lead author of a paper on this technique.

Joining Suh on the paper are co-lead author Tao Pang PhD ’23, a roboticist at Boston Dynamics AI Institute; Lujie Yang, an EECS graduate student; and senior author Russ Tedrake, the Toyota Professor of EECS, Aeronautics and Astronautics, and Mechanical Engineering, and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research appears this week in IEEE Transactions on Robotics.

Learning about learning

Reinforcement learning is a machine-learning technique where an agent, like a robot, learns to complete a task through trial and error with a reward for getting closer to a goal. Researchers say this type of learning takes a black-box approach because the system must learn everything about the world through trial and error.

It has been used effectively for contact-rich manipulation planning, where the robot seeks to learn the best way to move an object in a specified manner.

In these figures, a simulated robot performs three contact-rich manipulation tasks: in-hand manipulation of a ball, picking up a plate, and manipulating a pen into a specific orientation. Image: Courtesy of the researchers

But because there may be billions of potential contact points that a robot must reason about when determining how to use its fingers, hands, arms, and body to interact with an object, this trial-and-error approach requires a great deal of computation.

“Reinforcement learning may need to go through millions of years in simulation time to actually be able to learn a policy,” Suh adds.

On the other hand, if researchers specifically design a physics-based model using their knowledge of the system and the task they want the robot to accomplish, that model incorporates structure about this world that makes it more efficient.

Yet physics-based approaches aren’t as effective as reinforcement learning when it comes to contact-rich manipulation planning — Suh and Pang wondered why.

They conducted a detailed analysis and found that a technique known as smoothing enables reinforcement learning to perform so well.

Many of the decisions a robot could make when determining how to manipulate an object aren’t important in the grand scheme of things. For instance, each infinitesimal adjustment of one finger, whether or not it results in contact with the object, doesn’t matter very much.  Smoothing averages away many of those unimportant, intermediate decisions, leaving a few important ones.

Reinforcement learning performs smoothing implicitly by trying many contact points and then computing a weighted average of the results. Drawing on this insight, the MIT researchers designed a simple model that performs a similar type of smoothing, enabling it to focus on core robot-object interactions and predict long-term behavior. They showed that this approach could be just as effective as reinforcement learning at generating complex plans.

“If you know a bit more about your problem, you can design more efficient algorithms,” Pang says.

A winning combination

Even though smoothing greatly simplifies the decisions, searching through the remaining decisions can still be a difficult problem. So, the researchers combined their model with an algorithm that can rapidly and efficiently search through all possible decisions the robot could make.

With this combination, the computation time was cut down to about a minute on a standard laptop.

They first tested their approach in simulations where robotic hands were given tasks like moving a pen to a desired configuration, opening a door, or picking up a plate. In each instance, their model-based approach achieved the same performance as reinforcement learning, but in a fraction of the time. They saw similar results when they tested their model in hardware on real robotic arms.

“The same ideas that enable whole-body manipulation also work for planning with dexterous, human-like hands. Previously, most researchers said that reinforcement learning was the only approach that scaled to dexterous hands, but Terry and Tao showed that by taking this key idea of (randomized) smoothing from reinforcement learning, they can make more traditional planning methods work extremely well, too,” Tedrake says.

However, the model they developed relies on a simpler approximation of the real world, so it cannot handle very dynamic motions, such as objects falling. While effective for slower manipulation tasks, their approach cannot create a plan that would enable a robot to toss a can into a trash bin, for instance. In the future, the researchers plan to enhance their technique so it could tackle these highly dynamic motions.

“If you study your models carefully and really understand the problem you are trying to solve, there are definitely some gains you can achieve. There are benefits to doing things that are beyond the black box,” Suh says.

This work is funded, in part, by Amazon, MIT Lincoln Laboratory, the National Science Foundation, and the Ocado Group.

• PAPER – Global Planning for Contact-Rich Manipulation via Local Smoothing of Quasi-dynamic Contact Models. Tao Pang, H.J. Terry Suh, Lujie Yang and Russ Tedrake.

Copyright: Brian Johnson

Researchers at the Max Planck Institute for Intelligent Systems and the University of Colorado Boulder have developed a soft shape display, a robot that can rapidly and precisely change its surface geometry to interact with objects and liquids, react to human touch, and display letters and numbers – all at the same time. The display demonstrates high performance applications and could appear in the future on the factory floor, in medical laboratories, or in your own home.

Imagine an iPad that’s more than just an iPad—with a surface that can morph and deform, allowing you to draw 3D designs, create haiku that jump out from the screen and even hold your partner’s hand from an ocean away.

That’s the vision of a team of engineers from the University of Colorado Boulder (CU Boulder) and the Max Planck Institute for Intelligent Systems (MPI-IS) in Stuttgart, Germany. In a new study published in Nature Communications, they’ve created a one-of-a-kind shape-shifting display that fits on a card table. The device is made from a 10-by-10 grid of soft robotic “muscles” that can sense outside pressure and pop up to create patterns. It’s precise enough to generate scrolling text and fast enough to shake a chemistry beaker filled with fluid.

It may also deliver something even rarer: the sense of touch in a digital age.

“As technology has progressed, we started with sending text over long distances, then audio and later video,” said Brian Johnson, one of two lead authors of the new study who earned his doctorate in mechanical engineering at CU Boulder in 2022 and is now a postdoctoral researcher at the Max Planck Institute for Intelligent Systems. “But we’re still missing touch.”

The innovation builds off a class of soft robots pioneered by a team led by Christoph Keplinger, formerly an assistant professor of mechanical engineering at CU Boulder and now a director at MPI-IS. They’re called Hydraulically Amplified Self-Healing ELectrostatic (HASEL) actuators. The prototype display isn’t ready for the market yet. But the researchers envision that, one day, similar technologies could lead to sensory gloves for virtual gaming or a smart conveyer belt that can undulate to sort apples from bananas.

“You could imagine arranging these sensing and actuating cells into any number of different shapes and combinations,” said Mantas Naris, co-lead author of the paper and a doctoral student in the Paul M. Rady Department of Mechanical Engineering. “There’s really no limit to what these technologies could, ultimately, lead to.”

Playing the accordion

The project has its origins in the search for a different kind of technology: synthetic organs.

In 2017, researchers led by Mark Rentschler, professor of mechanical engineering and biomedical engineering, secured funding from the National Science Foundation to develop what they call sTISSUE—squishy organs that behave and feel like real human body parts but are made entirely out of plastic-like materials.

“You could use these artificial organs to help develop medical devices or surgical robotic tools for much less cost than using real animal tissue,” said Rentschler, a co-author of the new study.

In developing that technology, however, the team landed on the idea of a tabletop display.

The group’s design is about the size of a Scrabble game board and, like one of those boards, is composed of small squares arranged in a grid. In this case, each one of the 100 squares is an individual HASEL actuator. The actuators are made of plastic pouches shaped like tiny accordions. If you pass an electric current through them, fluid shifts around inside the pouches, causing the accordion to expand and jump up.

The actuators also include soft, magnetic sensors that can detect when you poke them. That allows for some fun activities, said Johnson.

“Because the sensors are magnet-based, we can use a magnetic wand to draw on the surface of the display,” he said.

Hear that?

Other research teams have developed similar smart tablets, but the CU Boulder display is softer, takes up a lot less room and is much faster. Each of its robotic muscles can move up to 3000 times per minute.

The researchers are focusing now on shrinking the actuators to increase the resolution of the display—almost like adding more pixels to a computer screen.

“Imagine if you could load an article onto your phone, and it renders as Braille on your screen,” Naris said.

The group is also working to flip the display inside out. That way, engineers could design a glove that pokes your fingertips, allowing you to “feel” objects in virtual reality.

And, Rentschler said, the display can bring something else: a little peace and quiet. “Our system is, essentially, silent. The actuators make almost no noise.”

Other CU Boulder co-authors of the new study include Nikolaus Correll, associate professor in the Department of Computer Science; Sean Humbert, professor of mechanical engineering; mechanical engineering graduate students Vani Sundaram, Angella Volchko and Khoi Ly; and alumni Shane Mitchell, Eric Acome and Nick Kellaris. Christoph Keplinger also served as a co-author in both of his roles at CU Boulder and MPI-IS.

• PAPER – A multifunctional soft robotic shape display with high-speed actuation, sensing, and control. Johnson, B.K., Naris, M., Sundaram, V., Volchko, A., Ly, K., Mitchell, S.K., Acome, E., Kellaris, N., Keplinger, C., Correll, N., Humbert, J.S., and Rentschler M.E. Nature Communications 14.1 (2023): 4516.

Image/Shutterstock.com

By Angharad Brewer Gillham, Frontiers science writer

Increasingly, social robots are being used for support in educational contexts. But does the sound of a social robot affect how well they perform, especially when dealing with teams of humans? Teamwork is a key factor in human creativity, boosting collaboration and new ideas. Danish scientists set out to understand whether robots using a voice designed to sound charismatic would be more successful as team creativity facilitators.

“We had a robot instruct teams of students in a creativity task. The robot either used a confident, passionate — ie charismatic — tone of voice or a normal, matter-of-fact tone of voice,” said Dr Kerstin Fischer of the University of Southern Denmark, corresponding author of the study in Frontiers in Communication. “We found that when the robot spoke in a charismatic speaking style, students’ ideas were more original and more elaborate.”

Can a robot be charismatic?

We know that social robots acting as facilitators can boost creativity, and that the success of facilitators is at least partly dependent on charisma: people respond to charismatic speech by becoming more confident and engaged. Fischer and her colleagues aimed to see if this effect could be reproduced with the voices of social robots by using a text-to-speech function engineered for characteristics associated with charismatic speaking, such as a specific pitch range and way of stressing words. Two voices were developed, one charismatic and one less expressive, based on a range of parameters which correlate with perceived speaker charisma.

The scientists recruited five classes of university students, all taking courses which included an element of team creativity. The students were told that they were testing a creativity workshop, which involved brainstorming ideas based on images and then using those ideas to come up with a new chocolate product. The workshop was led by videos of a robot speaking: introducing the task, reassuring the teams of students that there were no bad ideas, and then congratulating them for completing the task and asking them to fill out a self-evaluation questionnaire. The questionnaire evaluated the robot’s performance, the students’ own views on how their teamwork went, and the success of the session. The creativity of each session, as measured by the number of original ideas produced and how elaborate they were, was also measured by the researchers.

Powering creativity with charisma

The group that heard the charismatic voice rated the robot more positively, finding it more charismatic and interactive. Their perception of their teamwork was more positive, and they produced more original and elaborate ideas. They rated their teamwork more highly. However, the group that heard the non-charismatic voice perceived themselves as more resilient and efficient, possibly because a less charismatic leader led to better organization by the team members themselves, even though they produced fewer ideas.

“I had suspected that charismatic speech has very important effects, but our study provides clear evidence for the effect of charismatic speech on listener creativity,” said Dr Oliver Niebuhr of the University of Southern Denmark, co-author of the study. “This is the first time that such a link between charismatic voices, artificial speakers, and creativity outputs has been found.”

The scientists pointed out that although the sessions with the charismatic voice were generally more successful, not all the teams responded identically to the different voices: previous experiences in their different classes may have affected their response. Larger studies will be needed to understand how these external factors affected team performance.

“The robot was present only in videos, but one could suspect that more exposure or repeated exposure to the charismatic speaking style would have even stronger effects,” said Fischer. “Moreover, we have only varied a few features between the two robot conditions. We don’t know how the effect size would change if other or more features were varied. Finally, since charismatic speaking patterns differ between cultures, we would expect that the same stimuli will not yield the same results in all languages and cultures.”

• PAPER – Charismatic speech features in robot instructions enhance team creativity. Karen Fucinato, Oliver Niebuhr, Sladjana Nørskov, and Kerstin Fischer. Frontiers in Communication 8 (2023): 1115360.

CMU’s Autonomous Exploration Research Team has developed a suite of robotic systems and planners enabling robots to explore more quickly, probe the darkest corners of unknown environments, and create more accurate and detailed maps — all without human help.

By Aaron Aupperlee

A research group in Carnegie Mellon University’s Robotics Institute is creating the next generation of explorers — robots.

The Autonomous Exploration Research Team has developed a suite of robotic systems and planners enabling robots to explore more quickly, probe the darkest corners of unknown environments, and create more accurate and detailed maps. The systems allow robots to do all this autonomously, finding their way and creating a map without human intervention.

“You can set it in any environment, like a department store or a residential building after a disaster, and off it goes,” said Ji Zhang, a systems scientist in the Robotics Institute. “It builds the map in real-time, and while it explores, it figures out where it wants to go next. You can see everything on the map. You don’t even have to step into the space. Just let the robots explore and map the environment.”

The team has worked on exploration systems for more than three years. They’ve explored and mapped several underground mines, a parking garage, the Cohon University Center, and several other indoor and outdoor locations on the CMU campus. The system’s computers and sensors can be attached to nearly any robotic platform, transforming it into a modern-day explorer. The group uses a modified motorized wheelchair and drones for much of its testing.

Robots can explore in three modes using the group’s systems. In one mode, a person can control the robot’s movements and direction while autonomous systems keep it from crashing into walls, ceilings or other objects. In another mode, a person can select a point on a map and the robot will navigate to that point. The third mode is pure exploration. The robot sets off on its own, investigates the entire space and creates a map.

“This is a very flexible system to use in many applications, from delivery to search-and-rescue,” said Howie Choset, a professor in the Robotics Institute.

The group combined a 3D scanning lidar sensor, forward-looking camera and inertial measurement unit sensors with an exploration algorithm to enable the robot to know where it is, where it has been and where it should go next. The resulting systems are substantially more efficient than previous approaches, creating more complete maps while reducing the algorithm run time by half.

The new systems work in low-light, treacherous conditions where communication is spotty, like caves, tunnels and abandoned structures. A version of the group’s exploration system powered Team Explorer, an entry from CMU and Oregon State University in DARPA’s Subterranean Challenge. Team Explorer placed fourth in the final competition but won the Most Sectors Explored Award for mapping more of the route than any other team.

“All of our work is open-sourced. We are not holding anything back. We want to strengthen society with the capabilities of building autonomous exploration robots,” said Chao Cao, a Ph.D. student in robotics and the lead operator for Team Explorer. “It’s a fundamental capability. Once you have it, you can do a lot more.”

The group’s most recent work appeared in Science Robotics, which published “Representation Granularity Enables Time-Efficient Autonomous Exploration in Large, Complex Worlds” online. Past work has received top awards at prestigious robotics conferences. “TARE: A Hierarchical Framework for Efficiently Exploring Complex 3D Environments” won the Best Paper and Best Systems Paper awards at the Robotics Science and Systems Conference in 2021. It was the first time in the conference’s history that a paper received both awards. “FAR Planner: Fast, Attemptable Route Planner Using Dynamic Visibility Update” won the Best Student Paper Award at the International Conference on Intelligent Robots and Systems in 2022.

More information is available on the group’s website.

• PAPER – Representation granularity enables time-efficient autonomous exploration in large, complex worlds. C. Cao, H. Zhu, Z. Ren, H. Choset, and J. Zhang. Science Robotics, 8(80), eadf0970.

Scientists at the Max Planck Institute for Intelligent Systems in Stuttgart have developed a magnetically controlled soft medical robot with a unique, flexible structure inspired by the body of a pangolin. The robot is freely movable despite built-in hard metal components. Thus, depending on the magnetic field, it can adapt its shape to be able to move and can emit heat when needed, allowing for functionalities such as selective cargo transportation and release as well as mitigation of bleeding.

Pangolins are fascinating creatures. This animal looks like a walking pine cone, as it is the only mammal completely covered with hard scales. The scales are made of keratin, just like our hair and nails. The scales overlap and are directly connected to the underlying soft skin layer. This special arrangement allows the animals to curl up into a ball in case of danger.

While pangolins have many other unique features, researchers from the Physical Intelligence Department at the Max Planck Institute for Intelligent Systems in Stuttgart, which is led by Prof. Dr. Metin Sitti, were particularly fascinated by how pangolins can curl up their scale-covered bodies in a flash. They took the animal as a model and developed a flexible robot made of soft and hard components that, just like the animal, become a sphere in the blink of an eye – with the additional feature that the robot can emit heat when needed.

In a research paper published in Nature Communications, first author Ren Hao Soon and his colleagues present a robot design that is no more than two centimeters long and consists of two layers: a soft layer made of a polymer studded with small magnetic particles and a hard component made of metal elements arranged in overlapping layers. Thus, although the robot is made of solid metal components, it is still soft and flexible for use inside the human body.

Fig. 1 shows the pangolin-inspired untethered magnetic robot. A Conceptual illustration of the pangolin-inspired robot operating in the small intestine. Robot is actuated with a low-frequency magnetic field and heated remotely with a high-frequency magnetic field. The pangolin’s body consist of individual overlapping hard keratin scales. The robot inspired by this overlapping design is shown on the right. Images of pangolins used under Standard licence from Shutterstock.

When the robot is exposed to a low-frequency magnetic field, the researchers can roll up the robot and move it back and forth as they wish. The metal elements stick out like the animal’s scales, without hurting any surrounding tissue. Once it is rolled up, the robot can transport particles such as medicines. The vision is that such a small machine will one day travel through our digestive system, for example.

Double useful: freely movable and hot

When the robot is exposed to a high-frequency magnetic field, it heats up to over 70oC thanks to the built-in metal. Thermal energy is used in several medical procedures, such as treating thrombosis, stopping bleeding and removing tumor tissue. Untethered robots that can move freely, even though they are made of hard elements such as metal and can also emit heat, are rare. The pangolin robot is therefore considered promising for modern medicine. It could one day reach even the narrowest and most sensitive regions in the body in a minimally invasive and gentle way and emit heat as needed. That is a vision of the future. Already today, in a video, the researchers are showing how they can flexibly steer the robot through animal tissue and artificial organs.

• PAPER – Pangolin-inspired untethered magnetic robot for on-demand biomedical heating applications. Ren Hao Soon, Zhen Yin, Metin Alp Dogan, Nihal Olcay Dogan, Mehmet Efe Tiryaki, Alp Can Karacakol, Asli Aydin, Pouria Esmaeili-Dokht, and Metin Sitti. Nature Communications, 14(1), 3320.

By Hayley Dunning and Caroline Brogan

The prototype drone, called FireDrone, could be sent into burning buildings or woodland to assess hazards and provide crucial first-hand data from danger zones. The data would then be sent to first responders to help inform their emergency response.

The drone is made of a new thermal aerogel insulation material and houses an inbuilt cooling system to help it withstand temperatures of up to 200°C for ten minutes. Currently at prototype stage, the researchers believe FireDrone could eventually be used to scope out fires for people and extra hazards to bolster firefighting.

Principal Investigator Professor Mirko Kovac, Director of the Aerial Robotics Lab at Imperial College London and Head of the Laboratory of Sustainability Robotics at Empa, said: “Until they enter the danger zone, firefighters can’t be certain of what or who they’ll find, and what challenges they’ll encounter.

“FireDrone could be sent in ahead to gather crucial information so that responders can prepare accordingly to keep themselves safe and potentially save more lives.”

– Professor Mirko Kovac

Animal-inspired trailblazers

Drones are already used from afar in firefighting to take aerial footage, hoist fire hoses up skyscrapers, or drop fire retardant in remote areas to slow the spread of wildfires. However, current drones developed for firefighting are unable to fly much closer lest their frames melt and their electronics fail.

Based on interviews with firefighters, the researchers knew drones that could get much closer could help to prepare first responders for entering burning buildings or woodland. Drones equipped with cameras and carbon dioxide (CO2) sensors, for example, could provide crucial information about the layout and composition of fires.

They looked to animals that live in extreme temperatures like the penguin, arctic fox, and spittlebug, for inspiration. All these have appropriate layers of fat, fur, or produce their own layers of thermoregulating material that allow them to thrive in extreme conditions.

Inside the drone, showing the layer of aerogel. Credit: Empa

To build the drone, they created a protective structural shell made of lightweight, thermally super-insulating materials like polyimide aerogel, and glass fibres. They coated this with super-reflecting aluminium to reflect heat. The super-insulation prevents the materials from shrinking and pore structures from degrading after exposure to high temperatures.

Within the protective exoskeleton, they placed the temperature-sensitive components, such as regular and infrared cameras, CO2 sensors, video transmitters, flight controllers, batteries, and radio receivers. They also used the release and evaporation of gas from the CO2 sensors to build a cooling system to keep temperatures down.

Temperature extremes

They tested the drone in temperature-controlled chambers and flew it close to flames at a firefighter training centre. They hope that their further work to miniaturise and add more sensors to the drone might lead to its deployment in real-life firefighting missions and help save lives.

FireDrone can also be used in extreme cold environments, in polar regions and in glaciers. The team has also tested the robot in a glacier tunnel in Switzerland to study how the system behaves in very cold temperatures.

FireDrone tested in a glacier

Although FireDrone is at prototype stage, the researchers say it is a step forward for the development of other drones that can withstand extreme temperatures and the team is now validating the technology with key industrial stakeholders and partners.

Professor Kovac said: “The application of drones is often limited by environmental factors like temperature. We demonstrate a way to overcome this and are convinced our findings will help to unleash the future power of drones for extreme environments”.

“Deploying robots in extreme environments provides great benefits to reducing risks to human lives, and who better to look to than animals that have evolved their own ways of adapting to these extremes using inspirating from how animals keep cool in heat.”

• PAPER – FireDrone: Multi-Environment Thermally Agnostic Aerial Robot.
  David Häusermann, Sam Bodry, Fabian Wiesemüller, Aslan Miriyev, Severin Siegrist, Fan Fu, Sabyasachi Gaan, Matthias M. Koebel, Wim J. Malfait, Shanyu Zhao, and Mirko Kovač. Advanced Intelligent Systems (2023): 2300101.

Dr. Volker Strobel, postdoctoral researcher; Prof. Marco Dorigo, research director of the F.R.S.-FNRS; and Alexandre Pacheco, doctoral student. The researchers from the Université Libre de Bruxelles, Belgium. Credit: IRIDIA, Université Libre de Bruxelles

In a new study, we demonstrate the potential of blockchain technology, known from cryptocurrencies such as Bitcoin and Ethereum, to secure the coordination of robot swarms. In experiments conducted with both real and simulated robots, we show how blockchain technology enables a robot swarm to neutralize harmful robots without human intervention, thus enabling the deployment of autonomous and safe robot swarms.

Robot swarms are multi-robot systems that consist of many robots that collaborate in order to perform a task. They do not need a central control unit but the collective behavior of the swarm is rather a result of local interactions among robots. Thanks to this decentralization, robot swarms can work independently of external infrastructure, such as the Internet. This makes them particularly suitable for applications in a wide range of different environments such as underground, underwater, at sea, and in space.

Even though current swarm robotics applications are exclusively demonstrated in research environments, experts anticipate that in the non-distant future, robot swarms will support us in our everyday life. Robot swarms might perform environmental monitoring, underwater exploration, infrastructure inspection, and waste management—and thus make significant contributions to the transition into a fossil-free future with low pollution and high quality of life. In some of these activities, robot swarms will even outperform humans, leading to higher-quality results while ensuring our safety.

Once robot swarms are deployed in the real world, however, it is very likely that some robots in a swarm will break down (for example, due to harsh weather conditions) or might even be hacked. Such robots will not behave as intended and are called “Byzantine” robots. Recent research has shown that the actions of a very small minority of such Byzantine robots in a swarm can—similar to a virus—spread in the swarm and thus break down the whole system. Although security issues are crucial for the real-world deployment of robot swarms, security research in swarm robotics is lacking behind.

In Internet networks, Byzantine users such as hackers, have been successfully prevented from manipulating information by using blockchain technology. Blockchain technology is the technology behind Bitcoin: it enables users to agree on `who owns what’ without requiring a trusted third party such as a bank. Originally, blockchain technology was only meant to exchange units of a digital currency, such as Bitcoin. However, some years after Bitcoin’s release, blockchain-based smart contracts were introduced by the Ethereum framework: these smart contracts are programming code executed in a blockchain network. As no one can manipulate or stop this code, smart contracts enable “code is law”: contracts are automatically executed and do not need a trusted third party, such as a court, to be enforced.

So far, it was not clear whether large robot swarms could be controlled using blockchain and smart contracts. To address this open question, we presented a comprehensive study with both real and simulated robots in a collective-sensing scenario: the goal of the robot swarm is to provide an estimate of an environmental feature. To do so the robots need to sample the environment and then agree on the feature value. In our experiments, each robot is a member of a blockchain network maintained by the robots themselves. The robots send their estimates of environmental features to a smart contract that is shared by all the robots in the swarm. These estimates are aggregated by the smart contract that uses them to generate the requested estimate of the environmental feature. In this smart contract, we implemented economic mechanisms that ensure that good (non-Byzantine) robots are rewarded for sending useful information, whereas harmful Byzantine robots are penalized. The resulting robot economy prevents the Byzantine robots from participating in the swarm activities and influencing the swarm behavior.

Adding a blockchain to a robot swarm increases the robots’ computational requirements, such as CPU, RAM, and disk space usage. In fact, it was an open question whether running blockchain software on real robot swarms was possible at all. Our experiments have demonstrated that this is indeed possible as the computational requirements are manageable: the additional CPU, RAM, and disk space usage have a minor impact on the robot performance. This successful integration of blockchain technology into robot swarms paves the way for a wide range of secure robotic applications. To favor these future developments, we have released our software frameworks as open-source.

• PAPER – Robot swarms neutralize harmful Byzantine robots using a blockchain-based token economy. Volker Strobel, Alexandre Pacheco, and Marco Dorigo. Science Robotics, 8(79), eabm4636.

MIT professor of materials science and engineering and brain and cognitive sciences Polina Anikeeva in her lab. Photo: Steph Stevens

By Jennifer Michalowski | McGovern Institute for Brain Research

MIT scientists have developed tiny, soft-bodied robots that can be controlled with a weak magnet. The robots, formed from rubbery magnetic spirals, can be programmed to walk, crawl, swim — all in response to a simple, easy-to-apply magnetic field.

“This is the first time this has been done, to be able to control three-dimensional locomotion of robots with a one-dimensional magnetic field,” says Professor Polina Anikeeva, whose team published an open-access paper on the magnetic robots in the journal Advanced Materials. “And because they are predominantly composed of polymer and polymers are soft, you don’t need a very large magnetic field to activate them. It’s actually a really tiny magnetic field that drives these robots,” adds Anikeeva, who is a professor of materials science and engineering and brain and cognitive sciences at MIT, a McGovern Institute for Brain Research associate investigator, as well as the associate director of MIT’s Research Laboratory of Electronics and director of MIT’s K. Lisa Yang Brain-Body Center.

The new robots are well suited to transport cargo through confined spaces and their rubber bodies are gentle on fragile environments, opening the possibility that the technology could be developed for biomedical applications. Anikeeva and her team have made their robots millimeters long, but she says the same approach could be used to produce much smaller robots.

Magnetically actuated fiber-based soft robots

Engineering magnetic robots

Anikeeva says that until now, magnetic robots have moved in response to moving magnetic fields. She explains that for these models, “if you want your robot to walk, your magnet walks with it. If you want it to rotate, you rotate your magnet.” That limits the settings in which such robots might be deployed. “If you are trying to operate in a really constrained environment, a moving magnet may not be the safest solution. You want to be able to have a stationary instrument that just applies magnetic field to the whole sample,” she explains.

Youngbin Lee PhD ’22, a former graduate student in Anikeeva’s lab, engineered a solution to this problem. The robots he developed in Anikeeva’s lab are not uniformly magnetized. Instead, they are strategically magnetized in different zones and directions so a single magnetic field can enable a movement-driving profile of magnetic forces.

Before they are magnetized, however, the flexible, lightweight bodies of the robots must be fabricated. Lee starts this process with two kinds of rubber, each with a different stiffness. These are sandwiched together, then heated and stretched into a long, thin fiber. Because of the two materials’ different properties, one of the rubbers retains its elasticity through this stretching process, but the other deforms and cannot return to its original size. So when the strain is released, one layer of the fiber contracts, tugging on the other side and pulling the whole thing into a tight coil. Anikeeva says the helical fiber is modeled after the twisty tendrils of a cucumber plant, which spiral when one layer of cells loses water and contracts faster than a second layer.

A third material — one whose particles have the potential to become magnetic — is incorporated in a channel that runs through the rubbery fiber. So once the spiral has been made, a magnetization pattern that enables a particular type of movement can be introduced.

“Youngbin thought very carefully about how to magnetize our robots to make them able to move just as he programmed them to move,” Anikeeva says. “He made calculations to determine how to establish such a profile of forces on it when we apply a magnetic field that it will actually start walking or crawling.”

To form a caterpillar-like crawling robot, for example, the helical fiber is shaped into gentle undulations, and then the body, head, and tail are magnetized so that a magnetic field applied perpendicular to the robot’s plane of motion will cause the body to compress. When the field is reduced to zero, the compression is released, and the crawling robot stretches. Together, these movements propel the robot forward. Another robot in which two foot-like helical fibers are connected with a joint is magnetized in a pattern that enables a movement more like walking.

Biomedical potential

This precise magnetization process generates a program for each robot and ensures that that once the robots are made, they are simple to control. A weak magnetic field activates each robot’s program and drives its particular type of movement. A single magnetic field can even send multiple robots moving in opposite directions, if they have been programmed to do so. The team found that one minor manipulation of the magnetic field has a useful effect: With the flip of a switch to reverse the field, a cargo-carrying robot can be made to gently shake and release its payload.

Anikeeva says she can imagine these soft-bodied robots — whose straightforward production will be easy to scale up — delivering materials through narrow pipes, or even inside the human body. For example, they might carry a drug through narrow blood vessels, releasing it exactly where it is needed. She says the magnetically-actuated devices have biomedical potential beyond robots as well, and might one day be incorporated into artificial muscles or materials that support tissue regeneration.

• PAPER – Magnetically Actuated Fiber-Based Soft Robots. Youngbin Lee, Florian Koehler, Tom Dillon, Gabriel Loke, Yoonho Kim, Juliette Marion, Marc-Joseph Antonini, Indie Garwood, Atharva Sahasrabudhe, Keisuke Nagao, Xuanhe Zhao, Yoel Fink, Ellen T. Roche, and Polina Anikeeva. Advanced Materials, 2301916.

Today’s robots are often static and isolated from humans in structured environments — you can think of robot arms employed by Amazon for picking and packaging products within warehouses. But the true potential of robotics lies in mobile robots operating alongside humans in messy environments like our homes and hospitals — this requires navigation skills.

Imagine dropping a robot in a completely unseen home and asking it to find an object, let’s say a toilet. Humans can do this effortlessly: when looking for a glass of water at a friend’s house we’re visiting for the first time, we can easily find the kitchen without going to bedrooms or storage closets. But teaching this kind of spatial common sense to robots is challenging.

Many learning-based visual navigation policies have been proposed to tackle this problem. But learned visual navigation policies have predominantly been evaluated in simulation. How well do different classes of methods work on a robot?

We present a large-scale empirical study of semantic visual navigation methods comparing representative methods from classical, modular, and end-to-end learning approaches across six homes with no prior experience, maps, or instrumentation. We find that modular learning works well in the real world, attaining a 90% success rate. In contrast, end-to-end learning does not, dropping from 77% simulation to 23% real-world success rate due to a large image domain gap between simulation and reality.

Object goal navigation

We instantiate semantic navigation with the Object Goal navigation task, where a robot starts in a completely unseen environment and is asked to find an instance of an object category, let’s say a toilet. The robot has access to only a first-person RGB and depth camera and a pose sensor.

This task is challenging. It requires not only spatial scene understanding of distinguishing free space and obstacles and semantic scene understanding of detecting objects, but also requires learning semantic exploration priors. For example, if a human wants to find a toilet in this scene, most of us would choose the hallway because it is most likely to lead to a toilet. Teaching this kind of common sense or semantic priors to an autonomous agent is challenging. While exploring the scene for the desired object, the robot also needs to remember explored and unexplored areas.

Methods

So how do we train autonomous agents capable of efficient navigation while tackling all these challenges? A classical approach to this problem builds a geometric map using depth sensors, explores the environment with a heuristic, like frontier exploration, which explores the closest unexplored region, and uses an analytical planner to reach exploration goals and the goal object as soon as it is in sight. An end-to-end learning approach predicts actions directly from raw observations with a deep neural network consisting of visual encoders for image frames followed by a recurrent layer for memory. A modular learning approach builds a semantic map by projecting predicted semantic segmentation using depth, predicts an exploration goal with a goal-oriented semantic policy as a function of the semantic map and the goal object, and reaches it with a planner.

Large-scale real-world empirical evaluation

While many approaches to navigate to objects have been proposed over the past few years, learned navigation policies have predominantly been evaluated in simulation, which opens the field to the risk of sim-only research that does not generalize to the real world. We address this issue through a large-scale empirical evaluation of representative classical, end-to-end learning, and modular learning approaches across 6 unseen homes and 6 goal object categories.

Results

We compare approaches in terms of success rate within a limited budget of 200 robot actions and Success weighted by Path Length (SPL), a measure of path efficiency. In simulation, all approaches perform comparably, at around 80% success rate. But in the real world, modular learning and classical approaches transfer really well, up from 81% to 90% and 78% to 80% success rates, respectively. While end-to-end learning fails to transfer, down from 77% to 23% success rate.

We illustrate these results qualitatively with one representative trajectory. All approaches start in a bedroom and are tasked with finding a couch. On the left, modular learning first successfully reaches the couch goal. In the middle, end-to-end learning fails after colliding too many times. On the right, the classical policy finally reaches the couch goal after a detour through the kitchen.

Result 1: modular learning is reliable

We find that modular learning is very reliable on a robot, with a 90% success rate. Here, we can see it finds a plant in a first home efficiently, a chair in a second home, and a toilet in a third.

Result 2: modular learning explores more efficiently than classical

Modular learning improves by 10% real-world success rate over the classical approach. On the left, the goal-oriented semantic exploration policy directly heads towards the bedroom and finds the bed in 98 steps with an SPL of 0.90. On the right, because frontier exploration is agnostic to the bed goal, the policy makes detours through the kitchen and the entrance hallway before finally reaching the bed in 152 steps with an SPL of 0.52. With a limited time budget, inefficient exploration can lead to failure.

Result 3: end-to-end learning fails to transfer

While classical and modular learning approaches work well on a robot, end-to-end learning does not, at only 23% success rate. The policy collides often, revisits the same places, and even fails to stop in front of goal objects when they are in sight.

Analysis

Insight 1: why does modular transfer while end-to-end does not?

Why does modular learning transfer so well while end-to-end learning does not? To answer this question, we reconstructed one real-world home in simulation and conducted experiments with identical episodes in sim and reality.

The semantic exploration policy of the modular learning approach takes a semantic map as input, while the end-to-end policy directly operates on the RGB-D frames. The semantic map space is invariant between sim and reality, while the image space exhibits a large domain gap. In this example, this gap leads to a segmentation model trained on real-world images to predict a bed false positive in the kitchen.

The semantic map domain invariance allows the modular learning approach to transfer well from sim to reality. In contrast, the image domain gap causes a large drop in performance when transferring a segmentation model trained in the real world to simulation and vice versa. If semantic segmentation transfers poorly from sim to reality, it is reasonable to expect an end-to-end semantic navigation policy trained on sim images to transfer poorly to real-world images.

Insight 2: sim vs real gap in error modes for modular learning

Surprisingly, modular learning works even better in reality than simulation. Detailed analysis reveals that a lot of the failures of the modular learning policy that occur in sim are due to reconstruction errors, which do not happen in reality. Visual reconstruction errors represent 10% out of the total 19% episode failures, and physical reconstruction errors another 5%. In contrast, failures in the real world are predominantly due to depth sensor errors, while most semantic navigation benchmarks in simulation assume perfect depth sensing. Besides explaining the performance gap between sim and reality for modular learning, this gap in error modes is concerning because it limits the usefulness of simulation to diagnose bottlenecks and further improve policies. We show representative examples of each error mode and propose concrete steps forward to close this gap in the paper.

Takeaways

For practitioners:

• Modular learning can reliably navigate to objects with 90% success.

For researchers:

• Models relying on RGB images are hard to transfer from sim to real => leverage modularity and abstraction in policies.
• Disconnect between sim and real error modes => evaluate semantic navigation on real robots.

For more content about robotics and machine learning, check out my blog.

• PAPER – Navigating to Objects in the Real World. Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Science Robotics, 8(79), eadf6991.

Image: shutterstock.com

By Inês Hipólito/Deborah Pirchner, Frontiers science writer

Inês Hipólito is a highly accomplished researcher, recognized for her work in esteemed journals and contributions as a co-editor. She has received research awards including the prestigious Talent Grant from the University of Amsterdam in 2021. After her PhD, she held positions at the Berlin School of Mind and Brain and Humboldt-Universität zu Berlin. Currently, she is a permanent lecturer of the philosophy of AI at Macquarie University, focusing on cognitive development and the interplay between augmented cognition (AI) and the sociocultural environment.

Inês co-leads a consortium project on ‘Exploring and Designing Urban Density. Neurourbanism as a Novel Approach in Global Health,’ funded by the Berlin University Alliance. She also serves as an ethicist of AI at Verses.

Beyond her research, she co-founded and serves as vice-president of the International Society for the Philosophy of the Sciences of the Mind. Inês is the host of the thought-provoking podcast ‘The PhilospHER’s Way’ and has actively contributed to the Women in Philosophy Committee and the Committee in Diversity and Inclusivity at the Australasian Association of Philosophy from 2017 to 2020.

As part of our Frontier Scientist series, Hipólito caught up with Frontiers to tell us about her career and research.

Image: Inês Hipólito

What inspired you to become a researcher?
Throughout my personal journey, my innate curiosity and passion for understanding our experience of the world have been the driving forces in my life. Interacting with inspiring teachers and mentors during my education further fueled my motivation to explore the possibilities of objective understanding. This led me to pursue a multidisciplinary path in philosophy and neuroscience, embracing the original intent of cognitive science for interdisciplinary collaboration. I believe that by bridging disciplinary gaps, we can gain an understanding of the human mind and its interaction with the world. This integrative approach enables me to contribute to both scientific knowledge and real-world applications benefitting individuals and society as a whole.

Can you tell us about the research you’re currently working on?
My research centers around cognitive development and its implications in the cognitive science of AI. Sociocultural contexts play a pivotal role in shaping cognitive development, ranging from fundamental cognitive processes to more advanced, semantically sophisticated cognitive activities that we acquire and engage with.

As our world becomes increasingly permeated by AI, my research focuses on two main aspects. Firstly, I investigate how smart environments such as online spaces, virtual reality, and digitalized citizenship influence context-dependent cognitive development. By exploring the impact of these environments, I aim to gain insights into how cognition is shaped and adapted within these technologically mediated contexts.

Secondly, I examine how AI design emerges from specific sociocultural settings. Rather than merely reflecting society, AI design embodies societal values and aspirations. I explore the intricate relationship between AI and its sociocultural origins to understand how technology can both shape and be influenced by the context in which it is developed.

In your opinion, why is your research important?
The aim of my work is to contribute to the understanding of the complex relationship between cognition and AI, focusing on the sociocultural dynamics that influence both cognitive development and the design of artificial intelligence systems. I am particularly interested in understanding and the paradoxical nature of AI development and its societal impact: while technology historically improved lives, AI has also brought attention to problematic biases and segregation highlighted in feminist technoscience literature.

As AI progresses, it is crucial to ensure that advancements benefit everyone and do not perpetuate historical inequalities. Inclusivity and equality should be prioritized, challenging dominant narratives that favor certain groups, particularly white males. Recognizing that AI technologies embody our implicit biases and reflect our attitudes towards diversity and our relationship with the natural world enables us to navigate the ethical and societal implications of AI more effectively.

Are there any common misconceptions about this area of research? How would you address them?
The common misconception of viewing the mind as a computer has significant implications for AI design and our understanding of cognition. When cognition is seen as a simple input-output process in the brain, it overlooks the embodied complexities of human experience and the biases embedded in AI design. This reductionist view fails to account for the importance of embodied interaction, cognitive development, mental health, well-being, and societal equity.

This subjective experience of the world cannot be reduced to mere information processing, as it is context-dependent and imbued with meanings partly constructed in societal power dynamics.

Because the environment is ever more AI-permeated, understanding how it is shaped by and shapes the human experience requires investigation beyond the conceiving of cognition as (meaningless) information processes. By recognizing the distributed and embodied nature of cognition, we can ensure that AI technologies are designed and integrated in a way that respects the complexities of human experience, embraces ambiguity, and promotes meaningful and equitable societal interactions.

What are some of the areas of research you’d like to see tackled in the years ahead?
In the years ahead, it is crucial to tackle several AI-related areas to shape a more inclusive and sustainable future:

Design AI to reduce bias and discrimination, ensuring equal opportunities for individuals from diverse backgrounds.

Make AI systems transparent and explainable, enabling people to understand how decisions are made and how to hold them accountable for unintended consequences.

Collaborate with diverse stakeholders to address biases, cultural sensitivities, and challenges faced by marginalized communities in AI development.

Consider the ecological impact, resource consumption, waste generation, and carbon footprint throughout the entire lifecycle of AI technologies.

How has open science benefited the reach and impact of your research?
Scientific knowledge that is publicly funded should be made freely available to align with the principles of open science. Open science emphasizes transparency, collaboration, and accessibility in scientific research and knowledge dissemination. By openly sharing AI-related knowledge, including code, data, and algorithms, we encourage diverse stakeholders to contribute their expertise, identify potential biases, and address ethical concerns within technoscience.

Furthermore, incorporating philosophical reasoning into the development of the philosophy of mind theory can inform ethical deliberation and decision-making in AI design and implementation by researchers and policymakers. This transparent and collaborative approach enables critical assessment and improvement of AI technologies to ensure fairness, diminishing of bias, and overall equity.

• PAPER – Enactive artificial intelligence: subverting gender norms in human-robot interaction. Hipólito, Inês, Katie Winkle, and Merete Lie. Frontiers in Neurorobotics 17 (2023): 1149303.

This article is republished from Frontiers in Robotics and AI blog. You can read the original article here.

Picture: Adobe Stock/shurkin_son

Artificial hands, even the most sophisticated prostheses, are still by far inferior to human hands. What they lack are the tactile abilities crucial for dexterity. Other challenges include linking sensing to action within the robotic system – and effectively linking it to the human user. Prof. Dr. Philipp Beckerle from FAU has joined with international colleagues to summarize the latest findings in this field of Robotics – and establish an agenda for future research. Their piece in the research journal Science Robotics suggests a sensorimotor control framework for haptically enabled robotic hands, inspired by principles of the human’s central nervous system. Their aim is to link tactile sensing to movement in human-centred, haptically enabled artificial hands. According to the European and American team of researchers, this approach promises improved dexterity for humans controlling robotic hands.

Tactile sensing needs to play a bigger role

“Human manual dexterity relies critically on touch”, explains Prof. Dr. Philipp Beckerle, head of FAU’s Chair of Autonomous Systems and Mechatronics (ASM). “Humans with intact motor function but insensate fingertips can find it very difficult to grasp or manipulate things.” This, he says, indicates that tactile sensing is necessary for human dexterity. “Bioinspired design suggests that lessons from human haptics could enhance the currently limited dexterity of artificial hands. But robotic and prosthetic hands make little use of the many tactile sensors nowadays available and are hence much less dexterous.”

Beckerle, a Mechatronics engineer, has just had the paper “A hierarchical sensorimotor control framework for human-in-the-loop robotic hands” published in the research journal Science Robotics. In this, he unfolds with international colleagues how advanced technologies now provide not only mechatronic and computational components for anthropomorphic limbs, but also sensing ones. The scientists therefore suggest that such recently developed tactile sensing technologies could be incorporated into a general concept of “electronic skins”. “These include dense arrays of normal-force-sensing tactile elements in contrast to fingertips with a more comprehensive force perception”, the paper reads. “This would provide a directional force-distribution map over the entire sensing surface, and complex three-dimensional architectures, mimicking the mechanical properties and multimodal sensing of human fingertips.” Tactile sensing systems mounted on mechatronic limbs could therefore provide robotic systems with the complex representations needed to characterize, identify and manipulate, e.g. objects.

Human principles as inspiration for future designs

To achieve haptically informed and dexterous machines, the researchers secondly propose taking inspiration from the principles of the hierarchically organised human central nervous system (CNS). The CNS controls, which signals the brain receives from tactile senses and sends back to the body. The authors propose a conceptual framework in which a bioinspired touch-enabled robot shares control with the human – to a degree that the human sets. Principals of the framework include parallel processing of tasks, integration of feedforward and feedback control as well as a dynamic balance between subconscious and conscious processing. These could not only be applied in the design of bionic limbs, but also that of virtual avatars or remotely navigated telerobots.

It remains yet another challenge though to effectively interface a human user with touch-enabled robotic hands. “Enhancing haptic robots with high-density tactile sensing can substantially improve their capabilities but raises questions about how best to transmit these signals to a human controller, how to navigate shared perception and action in human-machine systems”, the paper reads. It remains largely unclear how to manage agency and task assignment, to maximize utility and user experience in human-in-the-loop systems. “Particularly challenging is how to exploit the varied and abundant tactile data generated by haptic devices. Yet, human principles provide inspiration for the future design of mechatronic systems that can function like humans, alongside humans, and even as replacement parts for humans.”

Philipp Beckerle’s Chair is part of the FAU’s Departments of Electrical Engineering, Electronics and Information Technology as well as the Department of Artificial Intelligence in Biomedical Engineering. “Our mission at ASM is to research human-centric mechatronics and robotics and strive for solutions that combine the desired performance with user-friendly interaction properties”, Beckerle explains. “Our focus is on wearable systems such as prostheses or exoskeletons, cognitive systems such as collaborative or humanoid robots and generally on tasks with close human-robot interaction. The human factors are crucial in such scenarios in order to meet the user’s needs and to achieve synergetic interface as well as interaction between humans and machines.”

Apart from Prof. Dr. Beckerle, scientists from the Universities of Genoa, Pisa and Rome, Aalborg, Bangor and Pittsburgh as well as the Imperial College London and the University of Southern California, Los Angeles were contributing to the paper.

• PAPER – A hierarchical sensorimotor control framework for human-in-the-loop robotic hands. Lucia Seminara, Strahinja Dosen, Fulvio Mastrogiovanni, Matteo Bianchi, Simon Watt, Philipp Beckerle, Thrishantha Nanayakkara, Knut Drewing, Alessandro Moscatelli, Roberta L. Klatzky, and Gerald E. Loeb. Science Robotics, 8(78), eadd5434.

Artificial Pancreas System with Reinforcement Learning. Image credit: Harry Emerson

Scientists at the University of Bristol have shown that reinforcement learning, a type of machine learning in which a computer program learns to make decisions by trying different actions, significantly outperforms commercial blood glucose controllers in terms of safety and effectiveness. By using offline reinforcement learning, where the algorithm learns from patient records, the researchers improve on prior work, showing that good blood glucose control can be achieved by learning from the decisions of the patient rather than by trial and error.

Type 1 diabetes is one of the most prevalent auto-immune conditions in the UK and is characterised by an insufficiency of the hormone insulin, which is responsible for blood glucose regulation.

Many factors affect a person’s blood glucose and therefore it can be a challenging and burdensome task to select the correct insulin dose for a given scenario. Current artificial pancreas devices provide automated insulin dosing but are limited by their simplistic decision-making algorithms.

However a new study, published in the Journal of Biomedical Informatics, shows offline reinforcement learning could represent an important milestone of care for people living with the condition. The largest improvement was in children, who experienced an additional one-and-a-half hours in the target glucose range per day.

Children represent a particularly important group as they are often unable to manage their diabetes without assistance and an improvement of this size would result in markedly better long-term health outcomes.

Lead author Harry Emerson from Bristol’s Department of Engineering Mathematics, explained: “My research explores whether reinforcement learning could be used to develop safer and more effective insulin dosing strategies.

“These machine learning driven algorithms have demonstrated superhuman performance in playing chess and piloting self-driving cars, and therefore could feasibly learn to perform highly personalised insulin dosing from pre-collected blood glucose data.

“This particular piece of work focuses specifically on offline reinforcement learning, in which the algorithm learns to act by observing examples of good and bad blood glucose control.

“Prior reinforcement learning methods in this area predominantly utilise a process of trial-and-error to identify good actions, which could expose a real-world patient to unsafe insulin doses.”

Due to the high risk associated with incorrect insulin dosing, experiments were performed using the FDA-approved UVA/Padova simulator, which creates a suite of virtual patients to test type 1 diabetes control algorithms. State-of-the-art offline reinforcement learning algorithms were evaluated against one of the most widely used artificial pancreas control algorithms. This comparison was conducted across 30 virtual patients (adults, adolescents and children) and considered 7,000 days of data, with performance being evaluated in accordance with current clinical guidelines. The simulator was also extended to consider realistic implementation challenges, such as measurement errors, incorrect patient information and limited quantities of available data.

This work provides a basis for continued reinforcement learning research in glucose control; demonstrating the potential of the approach to improve the health outcomes of people with type 1 diabetes, while highlighting the method’s shortcomings and areas of necessary future development.

The researchers’ ultimate goal is to deploy reinforcement learning in real-world artificial pancreas systems. These devices operate with limited patient oversight and consequently will require significant evidence of safety and effectiveness to achieve regulatory approval.

Harry added: ”This research demonstrates machine learning’s potential to learn effective insulin dosing strategies from the pre-collected type 1 diabetes data. The explored method outperforms one of the most widely used commercial artificial pancreas algorithms and demonstrates an ability to leverage a person’s habits and schedule to respond more quickly to dangerous events.”

• PAPER – Offline reinforcement learning for safer blood glucose control in people with type 1 diabetes. Emerson, Harry, Matthew Guy, and Ryan McConville. Journal of Biomedical Informatics, 142, 104376.

MIT researchers developed a machine-learning technique that can autonomously drive a car or fly a plane through a very difficult “stabilize-avoid” scenario, in which the vehicle must stabilize its trajectory to arrive at and stay within some goal region, while avoiding obstacles. Image: Courtesy of the researchers

By Adam Zewe | MIT News Office

In the film “Top Gun: Maverick,” Maverick, played by Tom Cruise, is charged with training young pilots to complete a seemingly impossible mission — to fly their jets deep into a rocky canyon, staying so low to the ground they cannot be detected by radar, then rapidly climb out of the canyon at an extreme angle, avoiding the rock walls. Spoiler alert: With Maverick’s help, these human pilots accomplish their mission.

A machine, on the other hand, would struggle to complete the same pulse-pounding task. To an autonomous aircraft, for instance, the most straightforward path toward the target is in conflict with what the machine needs to do to avoid colliding with the canyon walls or staying undetected. Many existing AI methods aren’t able to overcome this conflict, known as the stabilize-avoid problem, and would be unable to reach their goal safely.

MIT researchers have developed a new technique that can solve complex stabilize-avoid problems better than other methods. Their machine-learning approach matches or exceeds the safety of existing methods while providing a tenfold increase in stability, meaning the agent reaches and remains stable within its goal region.

In an experiment that would make Maverick proud, their technique effectively piloted a simulated jet aircraft through a narrow corridor without crashing into the ground. 

“This has been a longstanding, challenging problem. A lot of people have looked at it but didn’t know how to handle such high-dimensional and complex dynamics,” says Chuchu Fan, the Wilson Assistant Professor of Aeronautics and Astronautics, a member of the Laboratory for Information and Decision Systems (LIDS), and senior author of a new paper on this technique.

Fan is joined by lead author Oswin So, a graduate student. The paper will be presented at the Robotics: Science and Systems conference.

The stabilize-avoid challenge

Many approaches tackle complex stabilize-avoid problems by simplifying the system so they can solve it with straightforward math, but the simplified results often don’t hold up to real-world dynamics.

More effective techniques use reinforcement learning, a machine-learning method where an agent learns by trial-and-error with a reward for behavior that gets it closer to a goal. But there are really two goals here — remain stable and avoid obstacles — and finding the right balance is tedious.

The MIT researchers broke the problem down into two steps. First, they reframe the stabilize-avoid problem as a constrained optimization problem. In this setup, solving the optimization enables the agent to reach and stabilize to its goal, meaning it stays within a certain region. By applying constraints, they ensure the agent avoids obstacles, So explains. 

Then for the second step, they reformulate that constrained optimization problem into a mathematical representation known as the epigraph form and solve it using a deep reinforcement learning algorithm. The epigraph form lets them bypass the difficulties other methods face when using reinforcement learning. 

“But deep reinforcement learning isn’t designed to solve the epigraph form of an optimization problem, so we couldn’t just plug it into our problem. We had to derive the mathematical expressions that work for our system. Once we had those new derivations, we combined them with some existing engineering tricks used by other methods,” So says.

No points for second place

To test their approach, they designed a number of control experiments with different initial conditions. For instance, in some simulations, the autonomous agent needs to reach and stay inside a goal region while making drastic maneuvers to avoid obstacles that are on a collision course with it.

This video shows how the researchers used their technique to effectively fly a simulated jet aircraft in a scenario where it had to stabilize to a target near the ground while maintaining a very low altitude and staying within a narrow flight corridor. Courtesy of the researchers.

When compared with several baselines, their approach was the only one that could stabilize all trajectories while maintaining safety. To push their method even further, they used it to fly a simulated jet aircraft in a scenario one might see in a “Top Gun” movie. The jet had to stabilize to a target near the ground while maintaining a very low altitude and staying within a narrow flight corridor.

This simulated jet model was open-sourced in 2018 and had been designed by flight control experts as a testing challenge. Could researchers create a scenario that their controller could not fly? But the model was so complicated it was difficult to work with, and it still couldn’t handle complex scenarios, Fan says.

The MIT researchers’ controller was able to prevent the jet from crashing or stalling while stabilizing to the goal far better than any of the baselines.

In the future, this technique could be a starting point for designing controllers for highly dynamic robots that must meet safety and stability requirements, like autonomous delivery drones. Or it could be implemented as part of larger system. Perhaps the algorithm is only activated when a car skids on a snowy road to help the driver safely navigate back to a stable trajectory.

Navigating extreme scenarios that a human wouldn’t be able to handle is where their approach really shines, So adds.

“We believe that a goal we should strive for as a field is to give reinforcement learning the safety and stability guarantees that we will need to provide us with assurance when we deploy these controllers on mission-critical systems. We think this is a promising first step toward achieving that goal,” he says.

Moving forward, the researchers want to enhance their technique so it is better able to take uncertainty into account when solving the optimization. They also want to investigate how well the algorithm works when deployed on hardware, since there will be mismatches between the dynamics of the model and those in the real world.

“Professor Fan’s team has improved reinforcement learning performance for dynamical systems where safety matters. Instead of just hitting a goal, they create controllers that ensure the system can reach its target safely and stay there indefinitely,” says Stanley Bak, an assistant professor in the Department of Computer Science at Stony Brook University, who was not involved with this research. “Their improved formulation allows the successful generation of safe controllers for complex scenarios, including a 17-state nonlinear jet aircraft model designed in part by researchers from the Air Force Research Lab (AFRL), which incorporates nonlinear differential equations with lift and drag tables.”

The work is funded, in part, by MIT Lincoln Laboratory under the Safety in Aerobatic Flight Regimes program.

• PAPER – Solving Stabilize-Avoid Optimal Control via Epigraph Form and Deep Reinforcement Learning. Jorn A. Cheney, Jonathan P. J. Stevenson, So, Oswin, and Chuchu Fan. arXiv preprint arXiv:2305.14154 (2023).

Researchers created “FluidLab,” a simulation environment with a diverse set of manipulation tasks involving complex fluid dynamics. Image: Alex Shipps/MIT CSAIL via Midjourney

Imagine you’re enjoying a picnic by a riverbank on a windy day. A gust of wind accidentally catches your paper napkin and lands on the water’s surface, quickly drifting away from you. You grab a nearby stick and carefully agitate the water to retrieve it, creating a series of small waves. These waves eventually push the napkin back toward the shore, so you grab it. In this scenario, the water acts as a medium for transmitting forces, enabling you to manipulate the position of the napkin without direct contact.

Humans regularly engage with various types of fluids in their daily lives, but doing so has been a formidable and elusive goal for current robotic systems. Hand you a latte? A robot can do that. Make it? That’s going to require a bit more nuance. 

FluidLab, a new simulation tool from researchers at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), enhances robot learning for complex fluid manipulation tasks like making latte art, ice cream, and even manipulating air. The virtual environment offers a versatile collection of intricate fluid handling challenges, involving both solids and liquids, and multiple fluids simultaneously. FluidLab supports modeling solid, liquid, and gas, including elastic, plastic, rigid objects, Newtonian and non-Newtonian liquids, and smoke and air. 

At the heart of FluidLab lies FluidEngine, an easy-to-use physics simulator capable of seamlessly calculating and simulating various materials and their interactions, all while harnessing the power of graphics processing units (GPUs) for faster processing. The engine is “differential,” meaning the simulator can incorporate physics knowledge for a more realistic physical world model, leading to more efficient learning and planning for robotic tasks. In contrast, most existing reinforcement learning methods lack that world model that just depends on trial and error. This enhanced capability, say the researchers, lets users experiment with robot learning algorithms and toy with the boundaries of current robotic manipulation abilities.

To set the stage, the researchers tested said robot learning algorithms using FluidLab, discovering and overcoming unique challenges in fluid systems. By developing clever optimization methods, they’ve been able to transfer these learnings from simulations to real-world scenarios effectively. 

“Imagine a future where a household robot effortlessly assists you with daily tasks, like making coffee, preparing breakfast, or cooking dinner. These tasks involve numerous fluid manipulation challenges. Our benchmark is a first step towards enabling robots to master these skills, benefiting households and workplaces alike,” says visiting researcher at MIT CSAIL and research scientist at the MIT-IBM Watson AI Lab Chuang Gan, the senior author on a new paper about the research. “For instance, these robots could reduce wait times and enhance customer experiences in busy coffee shops. FluidEngine is, to our knowledge, the first-of-its-kind physics engine that supports a wide range of materials and couplings while being fully differentiable. With our standardized fluid manipulation tasks, researchers can evaluate robot learning algorithms and push the boundaries of today’s robotic manipulation capabilities.”

Fluid fantasia

Over the past few decades, scientists in the robotic manipulation domain have mainly focused on manipulating rigid objects, or on very simplistic fluid manipulation tasks like pouring water. Studying these manipulation tasks involving fluids in the real world can also be an unsafe and costly endeavor. 

With fluid manipulation, it’s not always just about fluids, though. In many tasks, such as creating the perfect ice cream swirl, mixing solids into liquids, or paddling through the water to move objects, it’s a dance of interactions between fluids and various other materials. Simulation environments must support “coupling,” or how two different material properties interact. Fluid manipulation tasks usually require pretty fine-grained precision, with delicate interactions and handling of materials, setting them apart from straightforward tasks like pushing a block or opening a bottle. 

FluidLab’s simulator can quickly calculate how different materials interact with each other. 

Helping out the GPUs is “Taichi,” a domain-specific language embedded in Python. The system can compute gradients (rates of change in environment configurations with respect to the robot’s actions) for different material types and their interactions (couplings) with one another. This precise information can be used to fine-tune the robot’s movements for better performance. As a result, the simulator allows for faster and more efficient solutions, setting it apart from its counterparts.

The 10 tasks the team put forth fell into two categories: using fluids to manipulate hard-to-reach objects, and directly manipulating fluids for specific goals. Examples included separating liquids, guiding floating objects, transporting items with water jets, mixing liquids, creating latte art, shaping ice cream, and controlling air circulation. 

“The simulator works similarly to how humans use their mental models to predict the consequences of their actions and make informed decisions when manipulating fluids. This is a significant advantage of our simulator compared to others,” says Carnegie Mellon University PhD student Zhou Xian, another author on the paper. “While other simulators primarily support reinforcement learning, ours supports reinforcement learning and allows for more efficient optimization techniques. Utilizing the gradients provided by the simulator supports highly efficient policy search, making it a more versatile and effective tool.”

Next steps

FluidLab’s future looks bright. The current work attempted to transfer trajectories optimized in simulation to real-world tasks directly in an open-loop manner. For next steps, the team is working to develop a closed-loop policy in simulation that takes as input the state or the visual observations of the environments and performs fluid manipulation tasks in real time, and then transfers the learned policies in real-world scenes.

The platform is publicly publicly available, and researchers hope it will benefit future studies in developing better methods for solving complex fluid manipulation tasks.

“Humans interact with fluids in everyday tasks, including pouring and mixing liquids (coffee, yogurts, soups, batter), washing and cleaning with water, and more,” says University of Maryland computer science professor Ming Lin, who was not involved in the work. “For robots to assist humans and serve in similar capacities for day-to-day tasks, novel techniques for interacting and handling various liquids of different properties (e.g. viscosity and density of materials) would be needed and remains a major computational challenge for real-time autonomous systems. This work introduces the first comprehensive physics engine, FluidLab, to enable modeling of diverse, complex fluids and their coupling with other objects and dynamical systems in the environment. The mathematical formulation of ‘differentiable fluids’ as presented in the paper makes it possible for integrating versatile fluid simulation as a network layer in learning-based algorithms and neural network architectures for intelligent systems to operate in real-world applications.”

Gan and Xian wrote the paper alongside Hsiao-Yu Tung a postdoc in the MIT Department of Brain and Cognitive Sciences; Antonio Torralba, an MIT professor of electrical engineering and computer science and CSAIL principal investigator; Dartmouth College Assistant Professor Bo Zhu, Columbia University PhD student Zhenjia Xu, and CMU Assistant Professor Katerina Fragkiadaki. The team’s research is supported by the MIT-IBM Watson AI Lab, Sony AI, a DARPA Young Investigator Award, an NSF CAREER award, an AFOSR Young Investigator Award, DARPA Machine Common Sense, and the National Science Foundation.

The research was presented at the International Conference on Learning Representations earlier this month.

• PAPER – FluidLab: A Differentiable Environment for Benchmarking Complex Fluid Manipulation. Zhou Xian, Bo Zhu, Zhenjia Xu, Hsiao-Yu Tung, Antonio Torralba, Katerina Fragkiadaki, and Chuang Gan. arXiv preprint arXiv:2303.02346 (2023).

Most of the world is covered in oceans, which are unfortunately highly polluted. One of the strategies to combat the mounds of waste found in these very sensitive ecosystems – especially around coral reefs – is to employ robots to master the cleanup. However, existing underwater robots are mostly bulky with rigid bodies, unable to explore and sample in complex and unstructured environments, and are noisy due to electrical motors or hydraulic pumps. For a more suitable design, scientists at the Max Planck Institute for Intelligent Systems (MPI-IS) in Stuttgart looked to nature for inspiration. They configured a jellyfish-inspired, versatile, energy-efficient and nearly noise-free robot the size of a hand. Jellyfish-Bot is a collaboration between the Physical Intelligence and Robotic Materials departments at MPI-IS. “A Versatile Jellyfish-like Robotic Platform for Effective Underwater Propulsion and Manipulation” was published in Science Advances.

To build the robot, the team used electrohydraulic actuators through which electricity flows. The actuators serve as artificial muscles which power the robot. Surrounding these muscles are air cushions as well as soft and rigid components which stabilize the robot and make it waterproof. This way, the high voltage running through the actuators cannot contact the surrounding water. A power supply periodically provides electricity through thin wires, causing the muscles to contract and expand. This allows the robot to swim gracefully and to create swirls underneath its body.

“When a jellyfish swims upwards, it can trap objects along its path as it creates currents around its body. In this way, it can also collect nutrients. Our robot, too, circulates the water around it. This function is useful in collecting objects such as waste particles. It can then transport the litter to the surface, where it can later be recycled. It is also able to collect fragile biological samples such as fish eggs. Meanwhile, there is no negative impact on the surrounding environment. The interaction with aquatic species is gentle and nearly noise-free”, Tianlu Wang explains. He is a postdoc in the Physical Intelligence Department at MPI-IS and first author of the publication.

His co-author Hyeong-Joon Joo from the Robotic Materials Department continues: “70% of marine litter is estimated to sink to the seabed. Plastics make up more than 60% of this litter, taking hundreds of years to degrade. Therefore, we saw an urgent need to develop a robot to manipulate objects such as litter and transport it upwards. We hope that underwater robots could one day assist in cleaning up our oceans.”

Jellyfish-Bots are capable of moving and trapping objects without physical contact, operating either alone or with several in combination. Each robot works faster than other comparable inventions, reaching a speed of up to 6.1 cm/s. Moreover, Jellyfish-Bot only requires a low input power of around 100 mW. And it is safe for humans and fish should the polymer material insulating the robot one day be torn apart. Meanwhile, the noise from the robot cannot be distinguished from background levels. In this way Jellyfish-Bot interacts gently with its environment without disturbing it – much like its natural counterpart.

The robot consists of several layers: some stiffen the robot, others serve to keep it afloat or insulate it. A further polymer layer functions as a floating skin. Electrically powered artificial muscles known as HASELs are embedded into the middle of the different layers. HASELs are liquid dielectric-filled plastic pouches that are partially covered by electrodes. Applying a high voltage across an electrode charges it positively, while surrounding water is charged negatively. This generates a force between positively-charged electrode and negatively-charged water that pushes the oil inside the pouches back and forth, causing the pouches to contract and relax – resembling a real muscle. HASELs can sustain the high electrical stresses generated by the charged electrodes and are protected against water by an insulating layer. This is important, as HASEL muscles were never before used to build an underwater robot.

The first step was to develop Jellyfish-Bot with one electrode with six fingers or arms. In the second step, the team divided the single electrode into separated groups to independently actuate them.

“We achieved grasping objects by making four of the arms function as a propeller, and the other two as a gripper. Or we actuated only a subset of the arms, in order to steer the robot in different directions. We also looked into how we can operate a collective of several robots. For instance, we took two robots and let them pick up a mask, which is very difficult for a single robot alone. Two robots can also cooperate in carrying heavy loads. However, at this point, our Jellyfish-Bot needs a wire. This is a drawback if we really want to use it one day in the ocean”, Hyeong-Joon Joo says.

Perhaps wires powering robots will soon be a thing of the past. “We aim to develop wireless robots. Luckily, we have achieved the first step towards this goal. We have incorporated all the functional modules like the battery and wireless communication parts so as to enable future wireless manipulation”, Tianlu Wang continues. The team attached a buoyancy unit at the top of the robot and a battery and microcontroller to the bottom. They then took their invention for a swim in the pond of the Max Planck Stuttgart campus, and could successfully steer it along. So far, however, they could not direct the wireless robot to change course and swim the other way.

Knowing the team, it won’t take long to achieve this goal.

• PAPER – A versatile jellyfish-like robotic platform for effective underwater propulsion and manipulation. Wang, T., Joo, H. J., Song, S., Hu, W., Keplinger, C., and Sitti, M. Science Advances, 9(15), 2023, eadg0292.

Artificial muscles are a progressing technology that could one day enable robots to function like living organisms. Such muscles open up new possibilities for how robots can shape the world around us; from assistive wearable devices that can redefine our physical abilities at old age, to rescue robots that can navigate rubble in search of the missing. But just because artificial muscles can have a strong societal impact during use, doesn’t mean they have to leave a strong environmental impact after use.

The topic of sustainability in soft robotics has been brought into focus by an international team of researchers from the Max Planck Institute for Intelligent Systems (MPI-IS) in Stuttgart (Germany), the Johannes Kepler University (JKU) in Linz (Austria), and the University of Colorado (CU Boulder), Boulder (USA). The scientists collaborated to design a fully biodegradable, high performance artificial muscle – based on gelatin, oil, and bioplastics. They show the potential of this biodegradable technology by using it to animate a robotic gripper, which could be especially useful in single-use deployments such as for waste collection (watch the Youtube video). At the end of life, these artificial muscles can be disposed of in municipal compost bins; under monitored conditions, they fully biodegrade within six months.

We see an urgent need for sustainable materials in the accelerating field of soft robotics. Biodegradable parts could offer a sustainable solution especially for single-use applications, like for medical operations, search-and-rescue missions, and manipulation of hazardous substances. Instead of accumulating in landfills at the end of product life, the robots of the future could become compost for future plant growth,” says Ellen Rumley, a visiting scientist from CU Boulder working in the Robotic Materials Department at MPI-IS. Rumley is co-first author of the paper “Biodegradable electrohydraulic actuators for sustainable soft robots”, published in Science Advances.

Specifically, the team of researchers built an electrically driven artificial muscle called HASEL. In essence, HASELs are oil-filled plastic pouches that are partially covered by a pair of electrical conductors called electrodes. Applying a high voltage across the electrode pair causes opposing charges to build on them, generating a force between them that pushes oil to an electrode-free region of the pouch. This oil migration causes the pouch to contract, much like a real muscle. The key requirement for HASELs to deform is that the materials making up the plastic pouch and oil are electrical insulators, which can sustain the high electrical stresses generated by the charged electrodes.

One of the challenges for this project was to develop a conductive, soft, and fully biodegradable electrode. Researchers at Johannes Kepler University created a recipe based on a mixture of biopolymer gelatin and salts that can be directly cast onto HASEL actuators. “It was important for us to make electrodes suitable for these high-performance applications, but with readily available components and an accessible fabrication strategy. Since our presented formulation can be easily integrated in various types of electrically driven systems, it serves as a building block for future biodegradable applications,” states David Preninger, co-first author for this project and a scientist at the Soft Matter Physics Division at JKU.

The next step was finding suitable biodegradable plastics. Engineers for this type of materials are mainly concerned with properties like degradation rate or mechanical strength, not with electrical insulation; a requirement for HASELs that operate at a few thousand Volts. Nonetheless, some bioplastics showed good material compatibility with gelatin electrodes and sufficient electrical insulation. HASELs made from one specific material combination were even able to withstand 100,000 actuation cycles at several thousand Volts without signs of electrical failure or loss in performance. These biodegradable artificial muscles are electromechanically competitive with their non-biodegradable counterparts; an exciting result for promoting sustainability in artificial muscle technology.

“By showing the outstanding performance of this new materials system, we are giving an incentive for the robotics community to consider biodegradable materials as a viable material option for building robots”, Ellen Rumley continues. “The fact that we achieved such great results with bio-plastics hopefully also motivates other material scientists to create new materials with optimized electrical performance in mind.”

With green technology becoming ever more present, the team’s research project is an important step towards a paradigm shift in soft robotics. Using biodegradable materials for building artificial muscles is just one step towards paving a future for sustainable robotic technology.

• PAPER – Biodegradable electrohydraulic actuators for sustainable soft robots. Ellen H. Rumley, David Preninger, Alona Shagan Shomron, Philipp Rothemund, Florian Hartmann, Melanie Baumgartner, Nicholas Kellaris, Andreas Stojanovic, Zachary Yoder, Benjamin Karrer, Christoph Keplinger, and Martin Kaltenbrunner. Science Advances 9(12), 2023.

Credit: Thomas Hartung, Johns Hopkins University

By Liad Hollender, Frontiers science writer

Despite AI’s impressive track record, its computational power pales in comparison with that of the human brain. Scientists unveil a revolutionary path to drive computing forward: organoid intelligence (OI), where lab-grown brain organoids serve as biological hardware. “This new field of biocomputing promises unprecedented advances in computing speed, processing power, data efficiency, and storage capabilities – all with lower energy needs,” say the authors in an article published in Frontiers in Science.

Artificial intelligence (AI) has long been inspired by the human brain. This approach proved highly successful: AI boasts impressive achievements – from diagnosing medical conditions to composing poetry. Still, the original model continues to outperform machines in many ways. This is why, for example, we can ‘prove our humanity’ with trivial image tests online. What if instead of trying to make AI more brain-like, we went straight to the source?

Scientists across multiple disciplines are working to create revolutionary biocomputers where three-dimensional cultures of brain cells, called brain organoids, serve as biological hardware. They describe their roadmap for realizing this vision in the journal Frontiers in Science.

“We call this new interdisciplinary field ‘organoid intelligence’ (OI),” said Prof Thomas Hartung of Johns Hopkins University. “A community of top scientists has gathered to develop this technology, which we believe will launch a new era of fast, powerful, and efficient biocomputing.”

What are brain organoids, and why would they make good computers?

Brain organoids are a type of lab-grown cell-culture. Even though brain organoids aren’t ‘mini brains’, they share key aspects of brain function and structure such as neurons and other brain cells that are essential for cognitive functions like learning and memory. Also, whereas most cell cultures are flat, organoids have a three-dimensional structure. This increases the culture’s cell density 1,000-fold, meaning that neurons can form many more connections.

But even if brain organoids are a good imitation of brains, why would they make good computers? After all, aren’t computers smarter and faster than brains?

“While silicon-based computers are certainly better with numbers, brains are better at learning,” Hartung explained. “For example, AlphaGo [the AI that beat the world’s number one Go player in 2017] was trained on data from 160,000 games. A person would have to play five hours a day for more than 175 years to experience these many games.” 

Brains are not only superior learners, they are also more energy efficient. For instance, the amount of energy spent training AlphaGo is more than is needed to sustain an active adult for a decade.

“Brains also have an amazing capacity to store information, estimated at 2,500TB,” Hartung added. “We’re reaching the physical limits of silicon computers because we cannot pack more transistors into a tiny chip. But the brain is wired completely differently. It has about 100bn neurons linked through over connection points. It’s an enormous power difference compared to our current technology.”

What would organoid intelligence bio computers look like?

According to Hartung, current brain organoids need to be scaled-up for OI. “They are too small, each containing about 50,000 cells. For OI, we would need to increase this number to 10 million,” he explained.

In parallel, the authors are also developing technologies to communicate with the organoids: in other words, to send them information and read out what they’re ‘thinking’. The authors plan to adapt tools from various scientific disciplines, such as bioengineering and machine learning, as well as engineer new stimulation and recording devices.

“We developed a brain-computer interface device that is a kind of an EEG cap for organoids, which we presented in an article published last August. It is a flexible shell that is densely covered with tiny electrodes that can both pick up signals from the organoid, and transmit signals to it,” said Hartung.

The authors envision that eventually OI would integrate a wide range of stimulation and recording tools. These will orchestrate interactions across networks of interconnected organoids that implement more complex computations.

Organoid intelligence could help prevent and treat neurological conditions

OI’s promise goes beyond computing and into medicine. Thanks to a groundbreaking technique developed by Noble Laureates John Gurdon and Shinya Yamanaka, brain organoids can be produced from adult tissues. This means that scientists can develop personalized brain organoids from skin samples of patients suffering from neural disorders, such as Alzheimer’s disease. They can then run multiple tests to investigate how genetic factors, medicines, and toxins influence these conditions.

“With OI, we could study the cognitive aspects of neurological conditions as well,” Hartung said. “For example, we could compare memory formation in organoids derived from healthy people and from Alzheimer’s patients, and try to repair relative deficits. We could also use OI to test whether certain substances, such as pesticides, cause memory or learning problems.”

Taking ethical considerations into account

Creating human brain organoids that can learn, remember, and interact with their environment raises complex ethical questions. For example, could they develop consciousness, even in a rudimentary form? Could they experience pain or suffering? And what rights would people have concerning brain organoids made from their cells?

The authors are acutely aware of these issues. “A key part of our vision is to develop OI in an ethical and socially responsible manner,” Hartung said. “For this reason, we have partnered with ethicists from the very beginning to establish an ‘embedded ethics’ approach. All ethical issues will be continuously assessed by teams made up of scientists, ethicists, and the public, as the research evolves.”

How far are we from the first organoid intelligence?

Even though OI is still in its infancy, a recently-published study by one of the article’s co-authors – Dr Brett Kagan, Chief Scientific Officer at Cortical Labs – provides proof of concept. His team showed that a normal, flat brain cell culture can learn to play the video game Pong.

“Their team is already testing this with brain organoids,” Hartung added. “And I would say that replicating this experiment with organoids already fulfills the basic definition of OI. From here on, it’s just a matter of building the community, the tools, and the technologies to realize OI’s full potential,” he concluded.

Interview with Prof Thomas Hartung

Image: Prof Thomas Hartung

To learn more about this exciting new field, we interviewed the senior author of the article, Prof Thomas Hartung. He is the director of the Center for Alternatives to Animal Testing in Europe (CAAT-Europe), and a professor at Johns Hopkins University’s Bloomberg School of Public Health.

How do you define organoid intelligence?

Reproducing cognitive functions – such as learning and sensory processing – in a lab-grown human-brain model.

How did this idea emerge?

I’m a pharmacologist and toxicologist, so I’m interested in developing medicines and identifying substances that are dangerous to our health, specifically those that affect brain development and function. This requires testing – ideally in conditions that mimic a living brain. For that reason, producing cultures of human brain cells has been a longstanding aim in the field.

This goal was finally realized in 2006 thanks to a groundbreaking technique developed by John B. Gurdon and Shinya Yamanaka, who received a Nobel prize for this achievement in 2012. This method allowed us to generate brain cells from fully developed tissues, such as the skin. Soon after, we began mass producing three-dimensional cultures of brain cells called brain organoids.

People asked if the organoids were thinking, if they were conscious even. I said: “no, they are too tiny. And more importantly, they don’t have any input nor output, so what would they be thinking about?” But later I began wondering: what if we changed this? What if we gave the organoids information about their environment and the means to interact with it? That was the birth of organoid intelligence.

How would you know what an organoid is ‘thinking’ about?

We’re building tools that will enable us to communicate with the organoids – send input and receive output. For example, we developed a recording/stimulation device that looks like a mini EEG-cap that surrounds the organoid. We’ve also been working on feeding biological inputs to brain organoids, for instance, by connecting them to retinal organoids, which respond to light. Our partner and co-author Alysson Muotri at the University of San Diego is already testing this approach by producing systems that combine several organoids.

My dream is to form a channel of communication between an artificial intelligence program and an OI system that would allow the two to explore each other’s capabilities. I imagine that form will follow function – that the organoid will change and develop towards creating meaningful inputs. This is a bit of philosophy, but my expectation is that we’ll see a lot of surprises.

What uses do you envision for organoid intelligence?

In my opinion, there are three main areas. The first is fundamental neuroscience – to understand how the brain generates cognitive functions, such as learning and memory. Even though current brain organoids are still far from being what one might call intelligent, they could still have the machinery to support basic cognitive operations.

The second area is toxicology and pharmacology. Since we can now produce brain organoids from skin samples, we can study individual disease characteristics of patients. We already have brain-organoid lines from Alzheimer’s patients, for example. And even though these organoids were made from skin cells, we still see hallmarks of the disease in them.

Next, we would like to test if there are also differences in their memory function, and if so, if we could repair it. We can also test whether substances, such as pesticides, worsen cognitive deficits, or cause them in brain organoids produced from healthy subjects. This is a very exciting line of research, which I believe is nearly within reach.

The third area is computing. As we laid out in our article, considering the brain’s size, its computational power is simply unmatched. Just for comparison, a supercomputer finally surpassed the computational power of a single human brain in 2022. But it cost $600m and occupies 680 square meters [about twice the area of a tennis court].

We’re also reaching the limits of computing. Moore’s Law, which states that the number of transistors in a microchip doubles every two years, has held for 60 years. But soon we won’t be able to physically fit more transistors into a chip. A single neuron, on the other hand, can connect to up to 10,000 other neurons – this is a very different way of processing and storing information. Through OI, we hope that we’ll be able to leverage the brain’s computational principles to build computers differently.

How do you intend to tackle ethical issues that might arise from organoid intelligence?

There are many questions that we face now, ranging from the rights of people over organoids developed from their cells, to understanding whether OI is conscious. I find this aspect of the work fascinating, and I believe it’s a fantastic opportunity to investigate the physical manifestation of concepts like sentience and consciousness.

We teamed up with Jeffrey Kahn of the Bloomberg School of Public Health at Johns Hopkins University at the very beginning, asking him to lead the discussion around the ethics of neural systems. We have come up with two main strategies. One is called embedded ethics: we want ethicists to closely observe the work, take part in the planning, and raise points early on. The second part focuses on the public – we intend to share our work broadly and clearly as it advances. We want to know how people feel about this technology and define our research plan accordingly.

How far are we from the first organoid intelligence?

Even though OI is still in its infancy, past work shows that it’s possible. A study by one of our partners and co-authors – Brett Kagan of the Cortical Labs – is a recent example. His team showed that a standard brain cell culture can learn to play the video game Pong. They are already experimenting with brain organoids, and I would say that replicating this with organoids already fulfills what we call OI.

Still, we are a long way from realizing OI’s full potential. When it becomes a real tool, it will look very different from these first baby steps we are taking now. The important thing is that it’s a starting point. I see this like sequencing the first genes of the human genome project: the enabling technology is in our hands, and we’re bound to learn a lot on the way.

This post is a combination of the original articles published on the Frontiers in Robotics and AI blog. You can read the originals here and here.

• PAPER – Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Smirnova, L., Caffo, B.S., Gracias, D.H., Huang, Q., Morales Pantoja, I.E., Tang, B., Zack, D.J., Berlinicke, C.A., Boyd, J.L., Harris, T.D. and Johnson, E.C., Frontiers in Science, 2023.

Robot fish. Image credit: Tsam Lung You

The robot fish was fitted with a twisted and coiled polymer (TCP) to drive it forward, a light-weight low cost device that relies on temperature change to generate movement, which also limits its speed.

A TCP works by contracting like muscles when heated, converting the energy into mechanical motion. The TCP used in this work is warmed by Joule heating – the pass of current through an electrical conductor produces thermal energy and heats up the conductor. By minimising the distance between the TCP on one side of the robot fish and the spring on the other, this activates the fin at the rear, enabling the robot fish to reach new speeds. The undulating flapping of its rear fin was measured at a frequency of 2Hz, two waves per second. The frequency of the electric current is the same as the frequency of tail flap.

The findings, published at the 6th IEEE-RAS International Conference on Soft Robotics (RoboSoft 2023), provide a new route to raising the actuation – the action of causing a machine or device to operate – frequency of TCPs through thermomechanical design and shows the possibility of using TCPs at high frequency in aqueous environments.

Lead author Tsam Lung You from Bristol’s Department of Engineering Mathematics said: “Twisted and coiled polymer (TCP) actuator is a promising novel actuator, exhibiting attractive properties of light weight, low-cost high energy density and simple fabrication process.

“They can be made from very easily assessable materials such as a fishing line and they contract and provide linear actuation when heated up. However, because of the time needed for heat dissipation during the relaxation phase, this makes them slow.”

By optimising the structural design of the TCP-spring antagonistic muscle pair and bringing their anchor points closer together, it allowed the posterior fin to swing at a larger angle for the same amount of TCP actuation.

Antagonistic muscles. Image credit: Tsam Lung You

Although this requires greater force, TCP is a strong actuator with high work energy density, and is still able to drive the fin.

Until now, TCPs have been mostly used for applications such as wearable devices and robotic hands. This work opens up more areas of application where TCP can be used, such as marine robots for underwater exploration and monitoring.

Tsam Lung You added: “Our robotic fish swam at the fastest actuation frequency found in a real TCP application and also the highest locomotion speed of a TCP application so far.

“This is really exciting as it opens up more opportunities of TCP application in different areas.”

The team now plan to expand the scale and develop a knifefish-inspired TCP-driven ribbon fin robot that can swim agilely in water.

• PAPER – Robotic Fish driven by Twisted and Coiled Polymer Actuators at High Frequencies.. You, T. L., Rossiter, J. M., and Philamore, H. In 6th IEEE-RAS International Conference on Soft Robotics (RoboSoft).

Makram Chahine, a PhD student in electrical engineering and computer science and an MIT CSAIL affiliate, leads a drone used to test liquid neural networks. Photo: Mike Grimmett/MIT CSAIL

By Rachel Gordon | MIT CSAIL

In the vast, expansive skies where birds once ruled supreme, a new crop of aviators is taking flight. These pioneers of the air are not living creatures, but rather a product of deliberate innovation: drones. But these aren’t your typical flying bots, humming around like mechanical bees. Rather, they’re avian-inspired marvels that soar through the sky, guided by liquid neural networks to navigate ever-changing and unseen environments with precision and ease.

Inspired by the adaptable nature of organic brains, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have introduced a method for robust flight navigation agents to master vision-based fly-to-target tasks in intricate, unfamiliar environments. The liquid neural networks, which can continuously adapt to new data inputs, showed prowess in making reliable decisions in unknown domains like forests, urban landscapes, and environments with added noise, rotation, and occlusion. These adaptable models, which outperformed many state-of-the-art counterparts in navigation tasks, could enable potential real-world drone applications like search and rescue, delivery, and wildlife monitoring.

The researchers’ recent study, published in Science Robotics, details how this new breed of agents can adapt to significant distribution shifts, a long-standing challenge in the field. The team’s new class of machine-learning algorithms, however, captures the causal structure of tasks from high-dimensional, unstructured data, such as pixel inputs from a drone-mounted camera. These networks can then extract crucial aspects of a task (i.e., understand the task at hand) and ignore irrelevant features, allowing acquired navigation skills to transfer targets seamlessly to new environments.

Drones navigate unseen environments with liquid neural networks.

“We are thrilled by the immense potential of our learning-based control approach for robots, as it lays the groundwork for solving problems that arise when training in one environment and deploying in a completely distinct environment without additional training,” says Daniela Rus, CSAIL director and the Andrew (1956) and Erna Viterbi Professor of Electrical Engineering and Computer Science at MIT. “Our experiments demonstrate that we can effectively teach a drone to locate an object in a forest during summer, and then deploy the model in winter, with vastly different surroundings, or even in urban settings, with varied tasks such as seeking and following. This adaptability is made possible by the causal underpinnings of our solutions. These flexible algorithms could one day aid in decision-making based on data streams that change over time, such as medical diagnosis and autonomous driving applications.”

A daunting challenge was at the forefront: Do machine-learning systems understand the task they are given from data when flying drones to an unlabeled object? And, would they be able to transfer their learned skill and task to new environments with drastic changes in scenery, such as flying from a forest to an urban landscape? What’s more, unlike the remarkable abilities of our biological brains, deep learning systems struggle with capturing causality, frequently over-fitting their training data and failing to adapt to new environments or changing conditions. This is especially troubling for resource-limited embedded systems, like aerial drones, that need to traverse varied environments and respond to obstacles instantaneously. 

The liquid networks, in contrast, offer promising preliminary indications of their capacity to address this crucial weakness in deep learning systems. The team’s system was first trained on data collected by a human pilot, to see how they transferred learned navigation skills to new environments under drastic changes in scenery and conditions. Unlike traditional neural networks that only learn during the training phase, the liquid neural net’s parameters can change over time, making them not only interpretable, but more resilient to unexpected or noisy data. 

In a series of quadrotor closed-loop control experiments, the drones underwent range tests, stress tests, target rotation and occlusion, hiking with adversaries, triangular loops between objects, and dynamic target tracking. They tracked moving targets, and executed multi-step loops between objects in never-before-seen environments, surpassing performance of other cutting-edge counterparts. 

The team believes that the ability to learn from limited expert data and understand a given task while generalizing to new environments could make autonomous drone deployment more efficient, cost-effective, and reliable. Liquid neural networks, they noted, could enable autonomous air mobility drones to be used for environmental monitoring, package delivery, autonomous vehicles, and robotic assistants. 

“The experimental setup presented in our work tests the reasoning capabilities of various deep learning systems in controlled and straightforward scenarios,” says MIT CSAIL Research Affiliate Ramin Hasani. “There is still so much room left for future research and development on more complex reasoning challenges for AI systems in autonomous navigation applications, which has to be tested before we can safely deploy them in our society.”

“Robust learning and performance in out-of-distribution tasks and scenarios are some of the key problems that machine learning and autonomous robotic systems have to conquer to make further inroads in society-critical applications,” says Alessio Lomuscio, professor of AI safety in the Department of Computing at Imperial College London. “In this context, the performance of liquid neural networks, a novel brain-inspired paradigm developed by the authors at MIT, reported in this study is remarkable. If these results are confirmed in other experiments, the paradigm here developed will contribute to making AI and robotic systems more reliable, robust, and efficient.”

Clearly, the sky is no longer the limit, but rather a vast playground for the boundless possibilities of these airborne marvels. 

Hasani and PhD student Makram Chahine; Patrick Kao ’22, MEng ’22; and PhD student Aaron Ray SM ’21 wrote the paper with Ryan Shubert ’20, MEng ’22; MIT postdocs Mathias Lechner and Alexander Amini; and Daniela Rus.

This research was supported, in part, by Schmidt Futures, the U.S. Air Force Research Laboratory, the U.S. Air Force Artificial Intelligence Accelerator, and the Boeing Co.

• PAPER – Robust flight navigation out of distribution with liquid neural networks. Makram Chahine, Ramin Hasani, Patrick Kao, Aaron Ray, Ryan Shubert, Mathias Lechner, Alexander Amini, and Daniela Rus. Science Robotics, 8(77), eadc8892.

Commercial and industrial deployments of robot fleets: package delivery (top left), food delivery (bottom left), e-commerce order fulfillment at Ambi Robotics (top right), autonomous taxis at Waymo (bottom right).

In the last few years we have seen an exciting development in robotics and artificial intelligence: large fleets of robots have left the lab and entered the real world. Waymo, for example, has over 700 self-driving cars operating in Phoenix and San Francisco and is currently expanding to Los Angeles. Other industrial deployments of robot fleets include applications like e-commerce order fulfillment at Amazon and Ambi Robotics as well as food delivery at Nuro and Kiwibot.

Figure 1: “Interactive Fleet Learning” (IFL) refers to robot fleets in industry and academia that fall back on human teleoperators when necessary and continually learn from them over time.

These robots use recent advances in deep learning to operate autonomously in unstructured environments. By pooling data from all robots in the fleet, the entire fleet can efficiently learn from the experience of each individual robot. Furthermore, due to advances in cloud robotics, the fleet can offload data, memory, and computation (e.g., training of large models) to the cloud via the Internet. This approach is known as “Fleet Learning,” a term popularized by Elon Musk in 2016 press releases about Tesla Autopilot and used in press communications by Toyota Research Institute, Wayve AI, and others. A robot fleet is a modern analogue of a fleet of ships, where the word fleet has an etymology tracing back to flēot (‘ship’) and flēotan (‘float’) in Old English.

Data-driven approaches like fleet learning, however, face the problem of the “long tail”: the robots inevitably encounter new scenarios and edge cases that are not represented in the dataset. Naturally, we can’t expect the future to be the same as the past! How, then, can these robotics companies ensure sufficient reliability for their services?

One answer is to fall back on remote humans over the Internet, who can interactively take control and “tele-operate” the system when the robot policy is unreliable during task execution. Teleoperation has a rich history in robotics: the world’s first robots were teleoperated during WWII to handle radioactive materials, and the Telegarden pioneered robot control over the Internet in 1994. With continual learning, the human teleoperation data from these interventions can iteratively improve the robot policy and reduce the robots’ reliance on their human supervisors over time. Rather than a discrete jump to full robot autonomy, this strategy offers a continuous alternative that approaches full autonomy over time while simultaneously enabling reliability in robot systems today.

The use of human teleoperation as a fallback mechanism is increasingly popular in modern robotics companies: Waymo calls it “fleet response,” Zoox calls it “TeleGuidance,” and Amazon calls it “continual learning.” Last year, a software platform for remote driving called Phantom Auto was recognized by Time Magazine as one of their Top 10 Inventions of 2022. And just last month, John Deere acquired SparkAI, a startup that develops software for resolving edge cases with humans in the loop.

A remote human teleoperator at Phantom Auto, a software platform for enabling remote driving over the Internet.

Despite this growing trend in industry, however, there has been comparatively little focus on this topic in academia. As a result, robotics companies have had to rely on ad hoc solutions for determining when their robots should cede control. The closest analogue in academia is interactive imitation learning (IIL), a paradigm in which a robot intermittently cedes control to a human supervisor and learns from these interventions over time. There have been a number of IIL algorithms in recent years for the single-robot, single-human setting including DAgger and variants such as HG-DAgger, SafeDAgger, EnsembleDAgger, and ThriftyDAgger; nevertheless, when and how to switch between robot and human control is still an open problem. This is even less understood when the notion is generalized to robot fleets, with multiple robots and multiple human supervisors.

IFL Formalism and Algorithms

To this end, in a recent paper at the Conference on Robot Learning we introduced the paradigm of Interactive Fleet Learning (IFL), the first formalism in the literature for interactive learning with multiple robots and multiple humans. As we’ve seen that this phenomenon already occurs in industry, we can now use the phrase “interactive fleet learning” as unified terminology for robot fleet learning that falls back on human control, rather than keep track of the names of every individual corporate solution (“fleet response”, “TeleGuidance”, etc.). IFL scales up robot learning with four key components:

1. On-demand supervision. Since humans cannot effectively monitor the execution of multiple robots at once and are prone to fatigue, the allocation of robots to humans in IFL is automated by some allocation policy . Supervision is requested “on-demand” by the robots rather than placing the burden of continuous monitoring on the humans.
2. Fleet supervision. On-demand supervision enables effective allocation of limited human attention to large robot fleets. IFL allows the number of robots to significantly exceed the number of humans (e.g., by a factor of 10:1 or more).
3. Continual learning. Each robot in the fleet can learn from its own mistakes as well as the mistakes of the other robots, allowing the amount of required human supervision to taper off over time.
4. The Internet. Thanks to mature and ever-improving Internet technology, the human supervisors do not need to be physically present. Modern computer networks enable real-time remote teleoperation at vast distances.

In the Interactive Fleet Learning (IFL) paradigm, M humans are allocated to the robots that need the most help in a fleet of N robots (where N can be much larger than M). The robots share policy and learn from human interventions over time.

We assume that the robots share a common control policy and that the humans share a common control policy . We also assume that the robots operate in independent environments with identical state and action spaces (but not identical states). Unlike a robot swarm of typically low-cost robots that coordinate to achieve a common objective in a shared environment, a robot fleet simultaneously executes a shared policy in distinct parallel environments (e.g., different bins on an assembly line).

The goal in IFL is to find an optimal supervisor allocation policy , a mapping from (the state of all robots at time t) and the shared policy to a binary matrix that indicates which human will be assigned to which robot at time t. The IFL objective is a novel metric we call the “return on human effort” (ROHE):

where the numerator is the total reward across robots and timesteps and the denominator is the total amount of human actions across robots and timesteps. Intuitively, the ROHE measures the performance of the fleet normalized by the total human supervision required. See the paper for more of the mathematical details.

Using this formalism, we can now instantiate and compare IFL algorithms (i.e., allocation policies) in a principled way. We propose a family of IFL algorithms called Fleet-DAgger, where the policy learning algorithm is interactive imitation learning and each Fleet-DAgger algorithm is parameterized by a unique priority function that each robot in the fleet uses to assign itself a priority score. Similar to scheduling theory, higher priority robots are more likely to receive human attention. Fleet-DAgger is general enough to model a wide range of IFL algorithms, including IFL adaptations of existing single-robot, single-human IIL algorithms such as EnsembleDAgger and ThriftyDAgger. Note, however, that the IFL formalism isn’t limited to Fleet-DAgger: policy learning could be performed with a reinforcement learning algorithm like PPO, for instance.

IFL Benchmark and Experiments

To determine how to best allocate limited human attention to large robot fleets, we need to be able to empirically evaluate and compare different IFL algorithms. To this end, we introduce the IFL Benchmark, an open-source Python toolkit available on Github to facilitate the development and standardized evaluation of new IFL algorithms. We extend NVIDIA Isaac Gym, a highly optimized software library for end-to-end GPU-accelerated robot learning released in 2021, without which the simulation of hundreds or thousands of learning robots would be computationally intractable. Using the IFL Benchmark, we run large-scale simulation experiments with N = 100 robots, M = 10 algorithmic humans, 5 IFL algorithms, and 3 high-dimensional continuous control environments (Figure 1, left).

We also evaluate IFL algorithms in a real-world image-based block pushing task with N = 4 robot arms and M = 2 remote human teleoperators (Figure 1, right). The 4 arms belong to 2 bimanual ABB YuMi robots operating simultaneously in 2 separate labs about 1 kilometer apart, and remote humans in a third physical location perform teleoperation through a keyboard interface when requested. Each robot pushes a cube toward a unique goal position randomly sampled in the workspace; the goals are programmatically generated in the robots’ overhead image observations and automatically resampled when the previous goals are reached. Physical experiment results suggest trends that are approximately consistent with those observed in the benchmark environments.

Takeaways and Future Directions

To address the gap between the theory and practice of robot fleet learning as well as facilitate future research, we introduce new formalisms, algorithms, and benchmarks for Interactive Fleet Learning. Since IFL does not dictate a specific form or architecture for the shared robot control policy, it can be flexibly synthesized with other promising research directions. For instance, diffusion policies, recently demonstrated to gracefully handle multimodal data, can be used in IFL to allow heterogeneous human supervisor policies. Alternatively, multi-task language-conditioned Transformers like RT-1 and PerAct can be effective “data sponges” that enable the robots in the fleet to perform heterogeneous tasks despite sharing a single policy. The systems aspect of IFL is another compelling research direction: recent developments in cloud and fog robotics enable robot fleets to offload all supervisor allocation, model training, and crowdsourced teleoperation to centralized servers in the cloud with minimal network latency.

While Moravec’s Paradox has so far prevented robotics and embodied AI from fully enjoying the recent spectacular success that Large Language Models (LLMs) like GPT-4 have demonstrated, the “bitter lesson” of LLMs is that supervised learning at unprecedented scale is what ultimately leads to the emergent properties we observe. Since we don’t yet have a supply of robot control data nearly as plentiful as all the text and image data on the Internet, the IFL paradigm offers one path forward for scaling up supervised robot learning and deploying robot fleets reliably in today’s world.

This post is based on the paper “Fleet-DAgger: Interactive Robot Fleet Learning with Scalable Human Supervision” by Ryan Hoque, Lawrence Chen, Satvik Sharma, Karthik Dharmarajan, Brijen Thananjeyan, Pieter Abbeel, and Ken Goldberg, presented at the Conference on Robot Learning (CoRL) 2022. For more details, see the paper on arXiv, CoRL presentation video on YouTube, open-source codebase on Github, high-level summary on Twitter, and project website.

If you would like to cite this article, please use the following bibtex:

@article{ifl_blog,
    title={Interactive Fleet Learning},
    author={Hoque, Ryan},
    url={https://bair.berkeley.edu/blog/2023/04/06/ifl/},
    journal={Berkeley Artificial Intelligence Research Blog},
    year={2023} 
}

The robotic system is shown in an experimental hive © Artificial Life Lab/U. of Graz/Hiveopolis

By Celia Luterbacher

Honeybees are famously finicky when it comes to being studied. Research instruments and conditions and even unfamiliar smells can disrupt a colony’s behavior. Now, a joint research team from the Mobile Robotic Systems Group in EPFL’s School of Engineering and School of Computer and Communication Sciences and the Hiveopolis project at Austria’s University of Graz have developed a robotic system that can be unobtrusively built into the frame of a standard honeybee hive.

Composed of an array of thermal sensors and actuators, the system measures and modulates honeybee behavior through localized temperature variations.

“Many rules of bee society – from collective and individual interactions to raising a healthy brood – are regulated by temperature, so we leveraged that for this study,” explains EPFL PhD student Rafael Barmak, first author on a paper on the system recently published in Science Robotics. “The thermal sensors create a snapshot of the bees’ collective behavior, while the actuators allow us to influence their movement by modulating thermal fields.”

“Previous studies on the thermal behavior of honeybees in winter have relied on observing the bees or manipulating the outside temperature,” adds Martin Stefanec of the University of Graz. “Our robotic system enables us to change the temperature from within the cluster, emulating the heating behavior of core bees there, and allowing us to study how the winter cluster actively regulates its temperature.”

A ‘biohybrid superorganism’ to mitigate colony collapse

Bee colonies are challenging to study in winter since they are sensitive to cold, and opening their hives risks harming them in addition to influencing their behavior. But thanks to the researchers’ biocompatible robotic system, they were able to study three experimental hives, located at the Artificial Life Lab at the University of Graz, during winter and to control them remotely from EPFL. Inside the device, a central processor coordinated the sensors, sent commands to the actuators, and transmitted data to the scientists, demonstrating that the system could be used to study bees with no intrusion – or even cameras – required.

Mobile Robotic Systems Group head Francesco Mondada explains that one of the most important aspects of the system – which he calls a ‘biohybrid superorganism’ for its combination of robotics with a colony of individuals acting as a living entity – is its ability to simultaneously observe and influence bee behavior.

“By gathering data on the bees’ position and creating warmer areas in the hive, we were able to encourage them to move around in ways they would never normally do in nature during the winter, when they tend to huddle together to conserve energy. This gives us the possibility to act on behalf of a colony, for example by directing it toward a food source, or discouraging it from dividing into too-small groups, which can threaten its survival.”

The robotic system is shown in an experimental hive © MOBOTS / EPFL / Hiveopolis

The scientists were able to prolong the survival of a colony following the death of its queen by distributing heat energy via the actuators. The system’s ability to mitigate colony collapse could have implications for bee survivability, which has become a growing environmental and food security concern as the pollinators’ global populations have declined.

Never-before-seen behaviors

In addition to its potential to support colonies, the system has shed light on honeybee behaviors that have never been observed, opening new avenues in biological research.

“The local thermal stimuli produced by our system revealed previously unreported dynamics that are generating exciting new questions and hypotheses,” says EPFL postdoctoral researcher and corresponding author Rob Mills. “For example, currently, no model can explain why we were able to encourage the bees to cross some cold temperature ‘valleys’ within the hive.”

The researchers now plan to use the system to study bees in summertime, which is a critical period for raising young. In parallel, the Mobile Robotic Systems Group is exploring systems using vibrational pathways to interact with honeybees.

“The biological acceptance aspect of this work is critical: the fact that the bees accepted the integration of electronics into the hive gives our device great potential for different scientific or agricultural applications,” says Mondada.

This work was supported by the EU H2020 FET project HIVEOPOLIS (no. 824069), coordinated by Thomas Schmickl, and by the Field of Excellence COLIBRI (Complexity of Life in basic Research and Innovation) at the University of Graz.

• PAPER – A robotic honeycomb for interaction with a honeybee colony. R. Barmak and M. Stefanec, D. N. Hofstadler, L. Piotet, S. Schönwetter-Fuchs-Schistek, F. Mondada, T. Schmickl, and R. Mills. Science Robotics, vol 8, Issue 76. https://doi.org/10.1126/scirobotics.add7385

MIT researchers have developed resilient artificial muscles that can enable insect-scale aerial robots to effectively recover flight performance after suffering severe damage. Photo: Courtesy of the researchers

By Adam Zewe | MIT News Office

Bumblebees are clumsy fliers. It is estimated that a foraging bee bumps into a flower about once per second, which damages its wings over time. Yet despite having many tiny rips or holes in their wings, bumblebees can still fly.

Aerial robots, on the other hand, are not so resilient. Poke holes in the robot’s wing motors or chop off part of its propellor, and odds are pretty good it will be grounded.

Inspired by the hardiness of bumblebees, MIT researchers have developed repair techniques that enable a bug-sized aerial robot to sustain severe damage to the actuators, or artificial muscles, that power its wings — but to still fly effectively.

They optimized these artificial muscles so the robot can better isolate defects and overcome minor damage, like tiny holes in the actuator. In addition, they demonstrated a novel laser repair method that can help the robot recover from severe damage, such as a fire that scorches the device.

Using their techniques, a damaged robot could maintain flight-level performance after one of its artificial muscles was jabbed by 10 needles, and the actuator was still able to operate after a large hole was burnt into it. Their repair methods enabled a robot to keep flying even after the researchers cut off 20 percent of its wing tip.

This could make swarms of tiny robots better able to perform tasks in tough environments, like conducting a search mission through a collapsing building or dense forest.

“We spent a lot of time understanding the dynamics of soft, artificial muscles and, through both a new fabrication method and a new understanding, we can show a level of resilience to damage that is comparable to insects,” says Kevin Chen, the D. Reid Weedon, Jr. Assistant Professor in the Department of Electrical Engineering and Computer Science (EECS), the head of the Soft and Micro Robotics Laboratory in the Research Laboratory of Electronics (RLE), and the senior author of the paper on these latest advances. “We’re very excited about this. But the insects are still superior to us, in the sense that they can lose up to 40 percent of their wing and still fly. We still have some catch-up work to do.”

Chen wrote the paper with co-lead authors Suhan Kim and Yi-Hsuan Hsiao, who are EECS graduate students; Younghoon Lee, a postdoc; Weikun “Spencer” Zhu, a graduate student in the Department of Chemical Engineering; Zhijian Ren, an EECS graduate student; and Farnaz Niroui, the EE Landsman Career Development Assistant Professor of EECS at MIT and a member of the RLE. The article appeared in Science Robotics.

Robot repair techniques

Using the repair techniques developed by MIT researchers, this microrobot can still maintain flight-level performance even after the artificial muscles that power its wings were jabbed by 10 needles and 20 percent of one wing tip was cut off. Credit: Courtesy of the researchers.

The tiny, rectangular robots being developed in Chen’s lab are about the same size and shape as a microcassette tape, though one robot weighs barely more than a paper clip. Wings on each corner are powered by dielectric elastomer actuators (DEAs), which are soft artificial muscles that use mechanical forces to rapidly flap the wings. These artificial muscles are made from layers of elastomer that are sandwiched between two razor-thin electrodes and then rolled into a squishy tube. When voltage is applied to the DEA, the electrodes squeeze the elastomer, which flaps the wing.

But microscopic imperfections can cause sparks that burn the elastomer and cause the device to fail. About 15 years ago, researchers found they could prevent DEA failures from one tiny defect using a physical phenomenon known as self-clearing. In this process, applying high voltage to the DEA disconnects the local electrode around a small defect, isolating that failure from the rest of the electrode so the artificial muscle still works.

Chen and his collaborators employed this self-clearing process in their robot repair techniques.

First, they optimized the concentration of carbon nanotubes that comprise the electrodes in the DEA. Carbon nanotubes are super-strong but extremely tiny rolls of carbon. Having fewer carbon nanotubes in the electrode improves self-clearing, since it reaches higher temperatures and burns away more easily. But this also reduces the actuator’s power density.

“At a certain point, you will not be able to get enough energy out of the system, but we need a lot of energy and power to fly the robot. We had to find the optimal point between these two constraints — optimize the self-clearing property under the constraint that we still want the robot to fly,” Chen says.

However, even an optimized DEA will fail if it suffers from severe damage, like a large hole that lets too much air into the device.

Chen and his team used a laser to overcome major defects. They carefully cut along the outer contours of a large defect with a laser, which causes minor damage around the perimeter. Then, they can use self-clearing to burn off the slightly damaged electrode, isolating the larger defect.

“In a way, we are trying to do surgery on muscles. But if we don’t use enough power, then we can’t do enough damage to isolate the defect. On the other hand, if we use too much power, the laser will cause severe damage to the actuator that won’t be clearable,” Chen says.

The team soon realized that, when “operating” on such tiny devices, it is very difficult to observe the electrode to see if they had successfully isolated a defect. Drawing on previous work, they incorporated electroluminescent particles into the actuator. Now, if they see light shining, they know that part of the actuator is operational, but dark patches mean they successfully isolated those areas.

The new research could make swarms of tiny robots better able to perform tasks in tough environments, like conducting a search mission through a collapsing building or dense forest. Photo: Courtesy of the researchers

Flight test success

Once they had perfected their techniques, the researchers conducted tests with damaged actuators — some had been jabbed by many needles while other had holes burned into them. They measured how well the robot performed in flapping wing, take-off, and hovering experiments.

Even with damaged DEAs, the repair techniques enabled the robot to maintain its flight performance, with altitude, position, and attitude errors that deviated only very slightly from those of an undamaged robot. With laser surgery, a DEA that would have been broken beyond repair was able to recover 87 percent of its performance.

“I have to hand it to my two students, who did a lot of hard work when they were flying the robot. Flying the robot by itself is very hard, not to mention now that we are intentionally damaging it,” Chen says.

These repair techniques make the tiny robots much more robust, so Chen and his team are now working on teaching them new functions, like landing on flowers or flying in a swarm. They are also developing new control algorithms so the robots can fly better, teaching the robots to control their yaw angle so they can keep a constant heading, and enabling the robots to carry a tiny circuit, with the longer-term goal of carrying its own power source.

“This work is important because small flying robots — and flying insects! — are constantly colliding with their environment. Small gusts of wind can be huge problems for small insects and robots. Thus, we need methods to increase their resilience if we ever hope to be able to use robots like this in natural environments,” says Nick Gravish, an associate professor in the Department of Mechanical and Aerospace Engineering at the University of California at San Diego, who was not involved with this research. “This paper demonstrates how soft actuation and body mechanics can adapt to damage and I think is an impressive step forward.”

This work is funded, in part, by the National Science Foundation (NSF) and a MathWorks Fellowship.

• PAPER – Laser-assisted failure recovery for dielectric elastomer actuators in aerial robots. Suhan Kim, Yi-Hsuan Hsiao, Younghoon Lee, Weikun Zhu, Zhijian Ren, Farnaz Niroui, and Yufeng Chen. Science Robotics, 8(76), eadf4278.

A team of MIT engineers is designing a kit of universal robotic parts that an astronaut could easily mix and match to build different robot “species” to fit various missions on the moon. Credit: hexapod image courtesy of the researchers, edited by MIT News

By Jennifer Chu | MIT News Office

When astronauts begin to build a permanent base on the moon, as NASA plans to do in the coming years, they’ll need help. Robots could potentially do the heavy lifting by laying cables, deploying solar panels, erecting communications towers, and building habitats. But if each robot is designed for a specific action or task, a moon base could become overrun by a zoo of machines, each with its own unique parts and protocols.

To avoid a bottleneck of bots, a team of MIT engineers is designing a kit of universal robotic parts that an astronaut could easily mix and match to rapidly configure different robot “species” to fit various missions on the moon. Once a mission is completed, a robot can be disassembled and its parts used to configure a new robot to meet a different task.

The team calls the system WORMS, for the Walking Oligomeric Robotic Mobility System. The system’s parts include worm-inspired robotic limbs that an astronaut can easily snap onto a base, and that work together as a walking robot. Depending on the mission, parts can be configured to build, for instance, large “pack” bots capable of carrying heavy solar panels up a hill. The same parts could be reconfigured into six-legged spider bots that can be lowered into a lava tube to drill for frozen water.

“You could imagine a shed on the moon with shelves of worms,” says team leader George Lordos, a PhD candidate and graduate instructor in MIT’s Department of Aeronautics and Astronautics (AeroAstro), in reference to the independent, articulated robots that carry their own motors, sensors, computer, and battery. “Astronauts could go into the shed, pick the worms they need, along with the right shoes, body, sensors and tools, and they could snap everything together, then disassemble it to make a new one. The design is flexible, sustainable, and cost-effective.”

Lordos’ team has built and demonstrated a six-legged WORMS robot. Last week, they presented their results at IEEE’s Aerospace Conference, where they also received the conference’s Best Paper Award.

MIT team members include Michael J. Brown, Kir Latyshev, Aileen Liao, Sharmi Shah, Cesar Meza, Brooke Bensche, Cynthia Cao, Yang Chen, Alex S. Miller, Aditya Mehrotra, Jacob Rodriguez, Anna Mokkapati, Tomas Cantu, Katherina Sapozhnikov, Jessica Rutledge, David Trumper, Sangbae Kim, Olivier de Weck, Jeffrey Hoffman, along with Aleks Siemenn, Cormac O’Neill, Diego Rivero, Fiona Lin, Hanfei Cui, Isabella Golemme, John Zhang, Jolie Bercow, Prajwal Mahesh, Stephanie Howe, and Zeyad Al Awwad, as well as Chiara Rissola of Carnegie Mellon University and Wendell Chun of the University of Denver.

Animal instincts

WORMS was conceived in 2022 as an answer to NASA’s Breakthrough, Innovative and Game-changing (BIG) Idea Challenge — an annual competition for university students to design, develop, and demonstrate a game-changing idea. In 2022, NASA challenged students to develop robotic systems that can move across extreme terrain, without the use of wheels.

A team from MIT’s Space Resources Workshop took up the challenge, aiming specifically for a lunar robot design that could navigate the extreme terrain of the moon’s South Pole — a landscape that is marked by thick, fluffy dust; steep, rocky slopes; and deep lava tubes. The environment also hosts “permanently shadowed” regions that could contain frozen water, which, if accessible, would be essential for sustaining astronauts.

As they mulled over ways to navigate the moon’s polar terrain, the students took inspiration from animals. In their initial brainstorming, they noted certain animals could conceptually be suited to certain missions: A spider could drop down and explore a lava tube, a line of elephants could carry heavy equipment while supporting each other down a steep slope, and a goat, tethered to an ox, could help lead the larger animal up the side of a hill as it transports an array of solar panels.

“As we were thinking of these animal inspirations, we realized that one of the simplest animals, the worm, makes similar movements as an arm, or a leg, or a backbone, or a tail,” says deputy team leader and AeroAstro graduate student Michael Brown. “And then the lightbulb went off: We could build all these animal-inspired robots using worm-like appendages.’”

The research team in Killian Court at MIT. Credit: Courtesy of the researchers

Snap on, snap off

Lordos, who is of Greek descent, helped coin WORMS, and chose the letter “O” to stand for “oligomeric,” which in Greek signifies “a few parts.”

“Our idea was that, with just a few parts, combined in different ways, you could mix and match and get all these different robots,” says AeroAstro undergraduate Brooke Bensche.

The system’s main parts include the appendage, or worm, which can be attached to a body, or chassis, via a “universal interface block” that snaps the two parts together through a twist-and-lock mechanism. The parts can be disconnected with a small tool that releases the block’s spring-loaded pins.

Appendages and bodies can also snap into accessories such as a “shoe,” which the team engineered in the shape of a wok, and a LiDAR system that can map the surroundings to help a robot navigate.

“In future iterations we hope to add more snap-on sensors and tools, such as winches, balance sensors, and drills,” says AeroAstro undergraduate Jacob Rodriguez.

The team developed software that can be tailored to coordinate multiple appendages. As a proof of concept, the team built a six-legged robot about the size of a go-cart. In the lab, they showed that once assembled, the robot’s independent limbs worked to walk over level ground. The team also showed that they could quickly assemble and disassemble the robot in the field, on a desert site in California.

In its first generation, each WORMS appendage measures about 1 meter long and weighs about 20 pounds. In the moon’s gravity, which is about one-sixth that of Earth’s, each limb would weigh about 3 pounds, which an astronaut could easily handle to build or disassemble a robot in the field. The team has planned out the specs for a larger generation with longer and slightly heavier appendages. These bigger parts could be snapped together to build “pack” bots, capable of transporting heavy payloads.

“There are many buzz words that are used to describe effective systems for future space exploration: modular, reconfigurable, adaptable, flexible, cross-cutting, et cetera,” says Kevin Kempton, an engineer at NASA’s Langley Research Center, who served as a judge for the 2022 BIG Idea Challenge. “The MIT WORMS concept incorporates all these qualities and more.”

This research was supported, in part, by NASA, MIT, the Massachusetts Space Grant, the National Science Foundation, and the Fannie and John Hertz Foundation.

• PAPER – WORMS: Field-Reconfigurable Robots for Extreme Lunar Terrain. George Lordos, Michael J. Brown, Kir Latyshev, and Aileen Liao. Proceedings of IEEE Aerospace 2023.

Author: D.Farina. Credits: Istituto Italiano di Tecnologia – © IIT, all rights reserved

Researchers at Istituto Italiano di Tecnologia (IIT-Italian Institute of Technology) in Genova has realized a new soft robot inspired by the biology of earthworms,which is able to crawl thanks to soft actuators that elongate or squeeze, when air passes through them or is drawn out. The prototype has been described in the international journal Scientific Reports of the Nature Portfolio, and it is the starting point for developing devices for underground exploration, but also search and rescue operations in confined spaces and the exploration of other planets.

Nature offers many examples of animals, such as snakes, earthworms, snails, and caterpillars, which use both the flexibility of their bodies and the ability to generate physical travelling waves along the length of their body to move and explore different environments. Some of their movements are also similar to plant roots.

Taking inspiration from nature and, at the same time, revealing new biological phenomena while developing new technologies is the main goal of the BioInspired Soft robotics lab coordinated by Barbara Mazzolai, and this earthworm-like robot is the latest invention coming from her group.

The creation of earthworm-like robot was made possible through a thorough understanding and application of earthworm locomotion mechanics. They use alternating contractions of muscle layers to propel themselves both below and above the soil surface by generating retrograde peristaltic waves. The individual segments of their body (metameres) have a specific quantity of fluid that controls the internal pressure to exert forces, and perform independent, localized and variable movement patterns.

IIT researchers have studied the morphology of earthworms and have found a way to mimic their muscle movements, their constant volume coelomic chambers and the function of their bristle-like hairs (setae) by creating soft robotic solutions.

The team developed a peristaltic soft actuator (PSA) that implements the antagonistic muscle movements of earthworms; from a neutral position it elongates when air is pumped into it and compresses when air is extracted from it. The entire body of the robotic earthworm is made of five PSA modules in series, connected with interlinks. The current prototype is 45 cm long and weighs 605 grams.

Each actuator has an elastomeric skin that encapsulates a known amount of fluid, thus mimicking the constant volume of internal coelomic fluid in earthworms. The earthworm segment becomes shorter longitudinally and wider circumferentially and exerts radial forces as the longitudinal muscles of an individual constant volume chamber contract. Antagonistically, the segment becomes longer along the anterior–posterior axis and thinner circumferentially with the contraction of circumferential muscles, resulting in penetration forces along the axis.

Every single actuator demonstrates a maximum elongation of 10.97mm at 1 bar of positive pressure and a maximum compression of 11.13mm at 0.5 bar of negative pressure, unique in its ability to generate both longitudinal and radial forces in a single actuator module.

In order to propel the robot on a planar surface, small passive friction pads inspired by earthworms’ setae were attached to the ventral surface of the robot. The robot demonstrated improved locomotion with a speed of 1.35mm/s.

This study not only proposes a new method for developing a peristaltic earthworm-like soft robot but also provides a deeper understanding of locomotion from a bioinspired perspective in different environments. The potential applications for this technology are vast, including underground exploration, excavation, search and rescue operations in subterranean environments and the exploration of other planets. This bioinspired burrowing soft robot is a significant step forward in the field of soft robotics and opens the door for further advancements in the future.

• PAPER – An earthworm‑like modular soft robot for locomotion in multi‑terrain environments. Riddhi Das, Saravana Prashanth Murali Babu, Francesco Visentin, Stefano Palagi, and Barbara Mazzolai. Sci Rep 13, 1571 (2023). https://doi.org/10.1038/s41598-023-28873-w

MIT engineers are hoping to help doctors tailor treatments to patients’ specific heart form and function, with a custom robotic heart. The team has developed a procedure to 3D print a soft and flexible replica of a patient’s heart. Image: Melanie Gonick, MIT

By Jennifer Chu | MIT News Office

No two hearts beat alike. The size and shape of the the heart can vary from one person to the next. These differences can be particularly pronounced for people living with heart disease, as their hearts and major vessels work harder to overcome any compromised function.

MIT engineers are hoping to help doctors tailor treatments to patients’ specific heart form and function, with a custom robotic heart. The team has developed a procedure to 3D print a soft and flexible replica of a patient’s heart. They can then control the replica’s action to mimic that patient’s blood-pumping ability.

The procedure involves first converting medical images of a patient’s heart into a three-dimensional computer model, which the researchers can then 3D print using a polymer-based ink. The result is a soft, flexible shell in the exact shape of the patient’s own heart. The team can also use this approach to print a patient’s aorta — the major artery that carries blood out of the heart to the rest of the body.

To mimic the heart’s pumping action, the team has fabricated sleeves similar to blood pressure cuffs that wrap around a printed heart and aorta. The underside of each sleeve resembles precisely patterned bubble wrap. When the sleeve is connected to a pneumatic system, researchers can tune the outflowing air to rhythmically inflate the sleeve’s bubbles and contract the heart, mimicking its pumping action. 

The researchers can also inflate a separate sleeve surrounding a printed aorta to constrict the vessel. This constriction, they say, can be tuned to mimic aortic stenosis — a condition in which the aortic valve narrows, causing the heart to work harder to force blood through the body.

Doctors commonly treat aortic stenosis by surgically implanting a synthetic valve designed to widen the aorta’s natural valve. In the future, the team says that doctors could potentially use their new procedure to first print a patient’s heart and aorta, then implant a variety of valves into the printed model to see which design results in the best function and fit for that particular patient. The heart replicas could also be used by research labs and the medical device industry as realistic platforms for testing therapies for various types of heart disease.

“All hearts are different,” says Luca Rosalia, a graduate student in the MIT-Harvard Program in Health Sciences and Technology. “There are massive variations, especially when patients are sick. The advantage of our system is that we can recreate not just the form of a patient’s heart, but also its function in both physiology and disease.”

Rosalia and his colleagues report their results in a study appearing in Science Robotics. MIT co-authors include Caglar Ozturk, Debkalpa Goswami, Jean Bonnemain, Sophie Wang, and Ellen Roche, along with Benjamin Bonner of Massachusetts General Hospital, James Weaver of Harvard University, and Christopher Nguyen, Rishi Puri, and Samir Kapadia at the Cleveland Clinic in Ohio.

Print and pump

In January 2020, team members, led by mechanical engineering professor Ellen Roche, developed a “biorobotic hybrid heart” — a general replica of a heart, made from synthetic muscle containing small, inflatable cylinders, which they could control to mimic the contractions of a real beating heart.

Shortly after those efforts, the Covid-19 pandemic forced Roche’s lab, along with most others on campus, to temporarily close. Undeterred, Rosalia continued tweaking the heart-pumping design at home.

“I recreated the whole system in my dorm room that March,” Rosalia recalls.

Months later, the lab reopened, and the team continued where it left off, working to improve the control of the heart-pumping sleeve, which they tested in animal and computational models. They then expanded their approach to develop sleeves and heart replicas that are specific to individual patients. For this, they turned to 3D printing.

“There is a lot of interest in the medical field in using 3D printing technology to accurately recreate patient anatomy for use in preprocedural planning and training,” notes Wang, who is a vascular surgery resident at Beth Israel Deaconess Medical Center in Boston.

An inclusive design

In the new study, the team took advantage of 3D printing to produce custom replicas of actual patients’ hearts. They used a polymer-based ink that, once printed and cured, can squeeze and stretch, similarly to a real beating heart.

As their source material, the researchers used medical scans of 15 patients diagnosed with aortic stenosis. The team converted each patient’s images into a three-dimensional computer model of the patient’s left ventricle (the main pumping chamber of the heart) and aorta. They fed this model into a 3D printer to generate a soft, anatomically accurate shell of both the ventricle and vessel.

The action of the soft, robotic models can be controlled to mimic the patient’s blood-pumping ability. Image: Melanie Gonick, MIT

The team also fabricated sleeves to wrap around the printed forms. They tailored each sleeve’s pockets such that, when wrapped around their respective forms and connected to a small air pumping system, the sleeves could be tuned separately to realistically contract and constrict the printed models.

The researchers showed that for each model heart, they could accurately recreate the same heart-pumping pressures and flows that were previously measured in each respective patient.

“Being able to match the patients’ flows and pressures was very encouraging,” Roche says. “We’re not only printing the heart’s anatomy, but also replicating its mechanics and physiology. That’s the part that we get excited about.”

Going a step further, the team aimed to replicate some of the interventions that a handful of the patients underwent, to see whether the printed heart and vessel responded in the same way. Some patients had received valve implants designed to widen the aorta. Roche and her colleagues implanted similar valves in the printed aortas modeled after each patient. When they activated the printed heart to pump, they observed that the implanted valve produced similarly improved flows as in actual patients following their surgical implants.

Finally, the team used an actuated printed heart to compare implants of different sizes, to see which would result in the best fit and flow — something they envision clinicians could potentially do for their patients in the future.

“Patients would get their imaging done, which they do anyway, and we would use that to make this system, ideally within the day,” says co-author Nguyen. “Once it’s up and running, clinicians could test different valve types and sizes and see which works best, then use that to implant.”

Ultimately, Roche says the patient-specific replicas could help develop and identify ideal treatments for individuals with unique and challenging cardiac geometries.

“Designing inclusively for a large range of anatomies, and testing interventions across this range, may increase the addressable target population for minimally invasive procedures,” Roche says.

This research was supported, in part, by the National Science Foundation, the National Institutes of Health, and the National Heart Lung Blood Institute.

• PAPER – Soft robotic patient-specific hydrodynamic model of aortic stenosis and ventricular remodeling. Luca Rosalia, Caglar Ozturk, Debkalpa Goswami, Jean Bonnemain, Sophie X. Wang, Benjamin Bonner, James C. Weaver, Rishi Puri, Samir Kapadia, Christopher T. Nguyen, and Ellen T. Roche. Science Robotics, 8 (75).

By Jędrzej Orbik, Charles Sun, Coline Devin, Glen Berseth

Reinforcement learning provides a conceptual framework for autonomous agents to learn from experience, analogously to how one might train a pet with treats. But practical applications of reinforcement learning are often far from natural: instead of using RL to learn through trial and error by actually attempting the desired task, typical RL applications use a separate (usually simulated) training phase. For example, AlphaGo did not learn to play Go by competing against thousands of humans, but rather by playing against itself in simulation. While this kind of simulated training is appealing for games where the rules are perfectly known, applying this to real world domains such as robotics can require a range of complex approaches, such as the use of simulated data, or instrumenting real-world environments in various ways to make training feasible under laboratory conditions. Can we instead devise reinforcement learning systems for robots that allow them to learn directly “on-the-job”, while performing the task that they are required to do? In this blog post, we will discuss ReLMM, a system that we developed that learns to clean up a room directly with a real robot via continual learning.

We evaluate our method on different tasks that range in difficulty. The top-left task has uniform white blobs to pickup with no obstacles, while other rooms have objects of diverse shapes and colors, obstacles that increase navigation difficulty and obscure the objects and patterned rugs that make it difficult to see the objects against the ground.

To enable “on-the-job” training in the real world, the difficulty of collecting more experience is prohibitive. If we can make training in the real world easier, by making the data gathering process more autonomous without requiring human monitoring or intervention, we can further benefit from the simplicity of agents that learn from experience. In this work, we design an “on-the-job” mobile robot training system for cleaning by learning to grasp objects throughout different rooms.

Lesson 1: The Benefits of Modular Policies for Robots.

People are not born one day and performing job interviews the next. There are many levels of tasks people learn before they apply for a job as we start with the easier ones and build on them. In ReLMM, we make use of this concept by allowing robots to train common-reusable skills, such as grasping, by first encouraging the robot to prioritize training these skills before learning later skills, such as navigation. Learning in this fashion has two advantages for robotics. The first advantage is that when an agent focuses on learning a skill, it is more efficient at collecting data around the local state distribution for that skill.

That is shown in the figure above, where we evaluated the amount of prioritized grasping experience needed to result in efficient mobile manipulation training. The second advantage to a multi-level learning approach is that we can inspect the models trained for different tasks and ask them questions, such as, “can you grasp anything right now” which is helpful for navigation training that we describe next.

Training this multi-level policy was not only more efficient than learning both skills at the same time but it allowed for the grasping controller to inform the navigation policy. Having a model that estimates the uncertainty in its grasp success (Ours above) can be used to improve navigation exploration by skipping areas without graspable objects, in contrast to No Uncertainty Bonus which does not use this information. The model can also be used to relabel data during training so that in the unlucky case when the grasping model was unsuccessful trying to grasp an object within its reach, the grasping policy can still provide some signal by indicating that an object was there but the grasping policy has not yet learned how to grasp it. Moreover, learning modular models has engineering benefits. Modular training allows for reusing skills that are easier to learn and can enable building intelligent systems one piece at a time. This is beneficial for many reasons, including safety evaluation and understanding.

Lesson 2: Learning systems beat hand-coded systems, given time

Many robotics tasks that we see today can be solved to varying levels of success using hand-engineered controllers. For our room cleaning task, we designed a hand-engineered controller that locates objects using image clustering and turns towards the nearest detected object at each step. This expertly designed controller performs very well on the visually salient balled socks and takes reasonable paths around the obstacles but it can not learn an optimal path to collect the objects quickly, and it struggles with visually diverse rooms. As shown in video 3 below, the scripted policy gets distracted by the white patterned carpet while trying to locate more white objects to grasp.

1)
2)
3)
4)
We show a comparison between (1) our policy at the beginning of training (2) our policy at the end of training (3) the scripted policy. In (4) we can see the robot’s performance improve over time, and eventually exceed the scripted policy at quickly collecting the objects in the room.

Given we can use experts to code this hand-engineered controller, what is the purpose of learning? An important limitation of hand-engineered controllers is that they are tuned for a particular task, for example, grasping white objects. When diverse objects are introduced, which differ in color and shape, the original tuning may no longer be optimal. Rather than requiring further hand-engineering, our learning-based method is able to adapt itself to various tasks by collecting its own experience.

However, the most important lesson is that even if the hand-engineered controller is capable, the learning agent eventually surpasses it given enough time. This learning process is itself autonomous and takes place while the robot is performing its job, making it comparatively inexpensive. This shows the capability of learning agents, which can also be thought of as working out a general way to perform an “expert manual tuning” process for any kind of task. Learning systems have the ability to create the entire control algorithm for the robot, and are not limited to tuning a few parameters in a script. The key step in this work allows these real-world learning systems to autonomously collect the data needed to enable the success of learning methods.

This post is based on the paper “Fully Autonomous Real-World Reinforcement Learning with Applications to Mobile Manipulation”, presented at CoRL 2021. You can find more details in our paper, on our website and the on the video. We provide code to reproduce our experiments. We thank Sergey Levine for his valuable feedback on this blog post.

Researchers have come up with an innovative approach to building deformable underwater robots using simple repeating substructures. The team has demonstrated the new system in two different example configurations, one like an eel, pictured here in the MIT tow tank. Credit: Courtesy of the researchers

By David L. Chandler | MIT News Office

Underwater structures that can change their shapes dynamically, the way fish do, push through water much more efficiently than conventional rigid hulls. But constructing deformable devices that can change the curve of their body shapes while maintaining a smooth profile is a long and difficult process. MIT’s RoboTuna, for example, was composed of about 3,000 different parts and took about two years to design and build.

Now, researchers at MIT and their colleagues — including one from the original RoboTuna team — have come up with an innovative approach to building deformable underwater robots, using simple repeating substructures instead of unique components. The team has demonstrated the new system in two different example configurations, one like an eel and the other a wing-like hydrofoil. The principle itself, however, allows for virtually unlimited variations in form and scale, the researchers say.

The work is being reported in the journal Soft Robotics, in a paper by MIT research assistant Alfonso Parra Rubio, professors Michael Triantafyllou and Neil Gershenfeld, and six others.

Existing approaches to soft robotics for marine applications are generally made on small scales, while many useful real-world applications require devices on scales of meters. The new modular system the researchers propose could easily be extended to such sizes and beyond, without requiring the kind of retooling and redesign that would be needed to scale up current systems.

The deformable robots are made with lattice-like pieces, called voxels, that are low density and have high stiffness. The deformable robots are made with lattice-like pieces, called voxels, that are low density and have high stiffness. Credit: Courtesy of the researchers

“Scalability is a strong point for us,” says Parra Rubio. Given the low density and high stiffness of the lattice-like pieces, called voxels, that make up their system, he says, “we have more room to keep scaling up,” whereas most currently used technologies “rely on high-density materials facing drastic problems” in moving to larger sizes.

The individual voxels in the team’s experimental, proof-of-concept devices are mostly hollow structures made up of cast plastic pieces with narrow struts in complex shapes. The box-like shapes are load-bearing in one direction but soft in others, an unusual combination achieved by blending stiff and flexible components in different proportions.

“Treating soft versus hard robotics is a false dichotomy,” Parra Rubio says. “This is something in between, a new way to construct things.” Gershenfeld, head of MIT’s Center for Bits and Atoms, adds that “this is a third way that marries the best elements of both.”

“Smooth flexibility of the body surface allows us to implement flow control that can reduce drag and improve propulsive efficiency, resulting in substantial fuel saving,” says Triantafyllou, who is the Henry L. and Grace Doherty Professor in Ocean Science and Engineering, and was part of the RoboTuna team.

Credit: Courtesy of the researchers.

In one of the devices produced by the team, the voxels are attached end-to-end in a long row to form a meter-long, snake-like structure. The body is made up of four segments, each consisting of five voxels, with an actuator in the center that can pull a wire attached to each of the two voxels on either side, contracting them and causing the structure to bend. The whole structure of 20 units is then covered with a rib-like supporting structure, and then a tight-fitting waterproof neoprene skin. The researchers deployed the structure in an MIT tow tank to show its efficiency in the water, and demonstrated that it was indeed capable of generating forward thrust sufficient to propel itself forward using undulating motions.

“There have been many snake-like robots before,” Gershenfeld says. “But they’re generally made of bespoke components, as opposed to these simple building blocks that are scalable.”

For example, Parra Rubio says, a snake-like robot built by NASA was made up of thousands of unique pieces, whereas for this group’s snake, “we show that there are some 60 pieces.” And compared to the two years spent designing and building the MIT RoboTuna, this device was assembled in about two days, he says.

The individual voxels are mostly hollow structures made up of cast plastic pieces with narrow struts in complex shapes. Credit: Courtesy of the researchers

The other device they demonstrated is a wing-like shape, or hydrofoil, made up of an array of the same voxels but able to change its profile shape and therefore control the lift-to-drag ratio and other properties of the wing. Such wing-like shapes could be used for a variety of purposes, ranging from generating power from waves to helping to improve the efficiency of ship hulls — a pressing demand, as shipping is a significant source of carbon emissions.

The wing shape, unlike the snake, is covered in an array of scale-like overlapping tiles, designed to press down on each other to maintain a waterproof seal even as the wing changes its curvature. One possible application might be in some kind of addition to a ship’s hull profile that could reduce the formation of drag-inducing eddies and thus improve its overall efficiency, a possibility that the team is exploring with collaborators in the shipping industry.

The team also created a wing-like hydrofoil. Credit: Courtesy of the researchers

Ultimately, the concept might be applied to a whale-like submersible craft, using its morphable body shape to create propulsion. Such a craft that could evade bad weather by staying below the surface, but without the noise and turbulence of conventional propulsion. The concept could also be applied to parts of other vessels, such as racing yachts, where having a keel or a rudder that could curve gently during a turn instead of remaining straight could provide an extra edge. “Instead of being rigid or just having a flap, if you can actually curve the way fish do, you can morph your way around the turn much more efficiently,” Gershenfeld says.

The research team included Dixia Fan of the Westlake University in China; Benjamin Jenett SM ’15, PhD ’ 20 of Discrete Lattice Industries; Jose del Aguila Ferrandis, Amira Abdel-Rahman and David Preiss of MIT; and Filippos Tourlomousis of the Demokritos Research Center of Greece. The work was supported by the U.S. Army Research Lab, CBA Consortia funding, and the MIT Sea Grant Program.

• PAPER – Modular Morphing Lattices for Large-Scale Underwater Continuum Robotic Structures. Alfonso Parra Rubio, Dixia Fan, Benjamin Jenett, José del Águila Ferrandis, Filippos Tourlomousis, Amira Abdel-Rahman, David Preiss, Michael Triantafyllou, and Neil Gershenfeld. Soft Robotics (2023).

A photograph of an eBiobot prototype, lit with blue microLEDs. Remotely controlled miniature biological robots have many potential applications in medicine, sensing and environmental monitoring. Image courtesy of Yongdeok Kim

By Liz Ahlberg Touchstone

First, they walked. Then, they saw the light. Now, miniature biological robots have gained a new trick: remote control.

The hybrid “eBiobots” are the first to combine soft materials, living muscle and microelectronics, said researchers at the University of Illinois Urbana-Champaign, Northwestern University and collaborating institutions. They described their centimeter-scale biological machines in the journal Science Robotics.

“Integrating microelectronics allows the merger of the biological world and the electronics world, both with many advantages of their own, to now produce these electronic biobots and machines that could be useful for many medical, sensing and environmental applications in the future,” said study co-leader Rashid Bashir, an Illinois professor of bioengineering and dean of the Grainger College of Engineering.

Rashid Bashir. Photo by L. Brian Stauffer

Bashir’s group has pioneered the development of biobots, small biological robots powered by mouse muscle tissue grown on a soft 3D-printed polymer skeleton. They demonstrated walking biobots in 2012 and light-activated biobots in 2016. The light activation gave the researchers some control, but practical applications were limited by the question of how to deliver the light pulses to the biobots outside of a lab setting.

The answer to that question came from Northwestern University professor John A. Rogers, a pioneer in flexible bioelectronics, whose team helped integrate tiny wireless microelectronics and battery-free micro-LEDs. This allowed the researchers to remotely control the eBiobots.

“This unusual combination of technology and biology opens up vast opportunities in creating self-healing, learning, evolving, communicating and self-organizing engineered systems. We feel that it’s a very fertile ground for future research with specific potential applications in biomedicine and environmental monitoring,” said Rogers, a professor of materials science and engineering, biomedical engineering and neurological surgery at Northwestern University and director of the Querrey Simpson Institute for Bioelectronics.

Remote control steering allows the eBiobots to maneuver around obstacles, as shown in this composite image of a bipedal robot traversing a maze. Image courtesy of Yongdeok Kim

To give the biobots the freedom of movement required for practical applications, the researchers set out to eliminate bulky batteries and tethering wires. The eBiobots use a receiver coil to harvest power and provide a regulated output voltage to power the micro-LEDs, said co-first author Zhengwei Li, an assistant professor of biomedical engineering at the University of Houston.

The researchers can send a wireless signal to the eBiobots that prompts the LEDs to pulse. The LEDs stimulate the light-sensitive engineered muscle to contract, moving the polymer legs so that the machines “walk.” The micro-LEDs are so targeted that they can activate specific portions of muscle, making the eBiobot turn in a desired direction. See a video on YouTube.

The researchers used computational modeling to optimize the eBiobot design and component integration for robustness, speed and maneuverability. Illinois professor of mechanical sciences and engineering Mattia Gazzola led the simulation and design of the eBiobots. The iterative design and additive 3D printing of the scaffolds allowed for rapid cycles of experiments and performance improvement, said Gazzola and co-first author Xiaotian Zhang, a postdoctoral researcher in Gazzola’s lab.

The eBiobots are the first wireless bio-hybrid machines, combining biological tissue, microelectronics and 3D-printed soft polymers. Image courtesy of Yongdeok Kim

The design allows for possible future integration of additional microelectronics, such as chemical and biological sensors, or 3D-printed scaffold parts for functions like pushing or transporting things that the biobots encounter, said co-first author Youngdeok Kim, who completed the work as a graduate student at Illinois.

The integration of electronic sensors or biological neurons would allow the eBiobots to sense and respond to toxins in the environment, biomarkers for disease and more possibilities, the researchers said.

“In developing a first-ever hybrid bioelectronic robot, we are opening the door for a new paradigm of applications for health care innovation, such as in-situ biopsies and analysis, minimum invasive surgery or even cancer detection within the human body,” Li said.

The National Science Foundation and the National Institutes of Health supported this work.

• PAPER – Remote control of muscle-driven miniature robots with battery-free wireless optoelectronics. Yongdeok Kim, Yiyuan Yang, Xiaotian Zhang,Zhengwei Li, Abraham Vázquez-Guardado, Insu Park, Jiaojiao Wang, Andrew I. Efimov, Zhi Dou, Yue Wang, Junehu Park, Haiwen Luan, Xinchen Ni, Yun Seong Kim, Janice Baek, Joshua Jaehyung Park, Zhaoqian Xie, Hangbo Zhao, Mattia Gazzola, John A. Rogers, and Rashid Bashir. Science Robotics 8.74 (2023): eadd1053.

By Peter Rüegg

Ecologists are increasingly using traces of genetic material left behind by living organisms left behind in the environment, called environmental DNA (eDNA), to catalogue and monitor biodiversity. Based on these DNA traces, researchers can determine which species are present in a certain area.

Obtaining samples from water or soil is easy, but other habitats – such as the forest canopy – are difficult for researchers to access. As a result, many species remain untracked in poorly explored areas.

Researchers at ETH Zurich and the Swiss Federal Institute for Forest, Snow and Landscape Research WSL, and the company SPYGEN have partnered to develop a special drone that can autonomously collect samples on tree branches.

(Video: ETH Zürich)

How the drone collects material

The drone is equipped with adhesive strips. When the aircraft lands on a branch, material from the branch sticks to these strips. Researchers can then extract DNA in the lab, analyse it and assign it to genetic matches of the various organisms using database comparisons.

But not all branches are the same: they vary in terms of their thickness and elasticity. Branches also bend and rebound when a drone lands on them. Programming the aircraft in such a way that it can still approach a branch autonomously and remain stable on it long enough to take samples was a major challenge for the roboticists.

“Landing on branches requires complex control,” explains Stefano Mintchev, Professor of Environmental Robotics at ETH Zurich and WSL. Initially, the drone does not know how flexible a branch is, so the researchers fitted it with a force sensing cage. This allows the drone to measure this factor at the scene and incorporate it into its flight manoeuvre.

Scheme: DNA is extracted from the collected branch material, amplified, sequenced and the sequences found are compared with databases. This allows the species to be identified. (Graphic: Stefano Mintchev / ETH Zürich)

Preparing rainforest operations at Zoo Zurich

Researchers have tested their new device on seven tree species. In the samples, they found DNA from 21 distinct groups of organisms, or taxa, including birds, mammals and insects. “This is encouraging, because it shows that the collection technique works,“ says Mintchev, who co-​authored the study that has appeared in the journal Science Robotics.

The researchers now want to improve their drone further to get it ready for a competition in which the aim is to detect as many different species as possible across 100 hectares of rainforest in Singapore in 24 hours.

To test the drone’s efficiency under conditions similar to those it will experience at the competition, Mintchev and his team are currently working at the Zoo Zurich’s Masoala Rainforest. “Here we have the advantage of knowing which species are present, which will help us to better assess how thorough we are in capturing all eDNA traces with this technique or if we’re missing something,“ Mintchev says.

For this event, however, the collection device must become more efficient and mobilize faster. In the tests in Switzerland, the drone collected material from seven trees in three days; in Singapore, it must be able to fly to and collect samples from ten times as many trees in just one day.

Collecting samples in a natural rainforest, however, presents the researchers with even tougher challenges. Frequent rain washes eDNA off surfaces, while wind and clouds impede drone operation. “We are therefore very curious to see whether our sampling method will also prove itself under extreme conditions in the tropics,” Mintchev says.

• PAPER – Drone-​assisted collection of environmental DNA from tree branches for biodiversity monitoring. Aucone E, Kirchgeorg E, Valentini A, Pellissier L, Deiner K, and Mintchev S. Science Robotics, 18 January 2023. DOI 10.1126/scirobotics.add5762

Joey’s design. Image credit: TL Nguyen, A Blight, A Pickering, A Barber, GH Jackson-Mills, JH Boyle, R Richardson, M Dogar, N Cohen

By Mischa Dijkstra, Frontiers science writer

Researchers from the University of Leeds have developed the first mini-robot, called Joey, that can find its own way independently through networks of narrow pipes underground, to inspect any damage or leaks. Joeys are cheap to produce, smart, small, and light, and can move through pipes inclined at a slope or over slippery or muddy sediment at the bottom of the pipes. Future versions of Joey will operate in swarms, with their mobile base on a larger ‘mother’ robot Kanga, which will be equipped with arms and tools for repairs to the pipes.

Beneath our streets lies a maze of pipes, conduits for water, sewage, and gas. Regular inspection of these pipes for leaks, or repair, normally requires these to be dug up. The latter is not only onerous and expensive – with an estimated annual cost of £5.5bn in the UK alone – but causes disruption to traffic as well as nuisance to people living nearby, not to mention damage to the environment.

Now imagine a robot that can find its way through the narrowest of pipe networks and relay images of damage or obstructions to human operators. This isn’t a pipedream anymore, shows a study in Frontiers in Robotics and AI by a team of researchers from the University of Leeds.

“Here we present Joey – a new miniature robot – and show that Joeys can explore real pipe networks completely on their own, without even needing a camera to navigate,” said Dr Netta Cohen, a professor at the University of Leeds and the final author on the study.

Joey is the first to be able to navigate all by itself through mazes of pipes as narrow as 7.5 cm across. Weighing just 70 g, it’s small enough to fit in the palm of your hand.

Pipebots project

The present work forms part of the ‘Pipebots’ project of the universities of Sheffield, Bristol, Birmingham, and Leeds, in collaboration with UK utility companies and other international academic and industrial partners.

First author Dr Thanh Luan Nguyen, a postdoctoral scientist at the University of Leeds who developed Joey’s control algorithms (or ‘brain’), said: “Underground water and sewer networks are some of the least hospitable environments, not only for humans, but also for robots. Sat Nav is not accessible undergound. And Joeys are tiny, so have to function with very simple motors, sensors, and computers that take little space, while the small batteries must be able to operate for long enough.”

Joey moves on 3D-printed ‘wheel-legs’ that roll through straight sections and walk over small obstacles. It is equipped with a range of energy-efficient sensors that measure its distance to walls, junctions, and corners, navigational tools, a microphone, and a camera and ‘spot lights’ to film faults in the pipe network and save the images. The prototype cost only £300 to produce.

Mud and slippery slopes

The team showed that Joey is able to find its way, without any instructions from human operators, through an experimental network of pipes including a T-junction, a left and right corner, a dead-end, an obstacle, and three straight sections. On average, Joey managed to explore about one meter of pipe network in just over 45 seconds.

To make life more difficult for the robot, the researchers verified that the robot easily moves up and down inclined pipes with realistic slopes. And to test Joey’s ability to navigate through muddy or slippery tubes, they also added sand and gooey gel (actually dishwashing liquid) to the pipes – again with success.

Importantly, the sensors are enough to allow Joey to navigate without the need to turn on the camera or use power-hungry computer vision. This saves energy and extends Joey’s current battery life. Whenever the battery runs low, Joey will return to its point of origin, to ‘feed’ on power.

Currently, Joeys have one weakness: they can’t right themselves if they inadvertently turn on their back, like an upside-down tortoise. The authors suggest that the next prototype will be able to overcome this challenge. Future generations of Joey should also be waterproof, to operate underwater in pipes entirely filled with liquid.

Joey’s future is collaborative

The Pipebots scientists aim to develop a swarm of Joeys that communicate and work together, based off a larger ‘mother’ robot named Kanga. Kanga, currently under development and testing by some of the same authors at Leeds School of Computing, will be equipped with more sophisticated sensors and repair tools such as robot arms, and carry multiple Joeys.

“Ultimately we hope to design a system that can inspect and map the condition of extensive pipe networks, monitor the pipes over time, and even execute some maintenance and repair tasks,” said Cohen.

“We envision the technology to scale up and diversify, creating an ecology of multi-species of robots that collaborate underground. In this scenario, groups of Joeys would be deployed by larger robots that have more power and capabilities but are restricted to the larger pipes. Meeting this challenge will require more research, development, and testing over 10 to 20 years. It may start to come into play around 2040 or 2050.” 

Top half: navigating through a T-junction in the pipe network. Bottom half: encountering an obstruction and turning back. Image credit: TL Nguyen, A Blight, A Pickering, A Barber, GH Jackson-Mills, JH Boyle, R Richardson, M Dogar, N Cohen

Top half: moving through sand, slippery goo, or mud. Bottom half: moving through pipe sloped at an angle. Image credit: TL Nguyen, A Blight, A Pickering, A Barber, GH Jackson-Mills, JH Boyle, R Richardson, M Dogar, N Cohen

• PAPER – Autonomous control for miniaturized mobile robots in unknown pipe networks. T. L. Nguyen, A. Blight, A. Pickering, G. Jackson-Mills, A. R. Barber, J. H. Boyle, R. Richardson, M. Dogar, and N. Cohen. Frontiers in Robotics and AI, 309.

Soft robots, or those made with materials like rubber, gels and cloth, have advantages over their harder, heavier counterparts, especially when it comes to tasks that require direct human interaction. Robots that could safely and gently help people with limited mobility grocery shop, prepare meals, get dressed, or even walk would undoubtedly be life-changing.

However, soft robots currently lack the strength needed to perform these sorts of tasks. This long-standing challenge — making soft robots stronger without compromising their ability to gently interact with their environment — has limited the development of these devices.

With the relationship between strength and softness in mind, a team of Penn Engineers has devised a new electrostatically controlled clutch which enables a soft robotic hand to be able to hold 4 pounds – about the weight of a bag of apples – which is 40 times more than the hand could lift without the clutch. In addition, the ability to perform this task requiring both a soft touch and strength was accomplished with only 125 volts of electricity, a third of the voltage required for current clutches.

Their safe, low-power approach could also enable wearable soft robotic devices that would simulate the sensation of holding a physical object in augmented- and virtual-reality environments.

James Pikul, Assistant Professor in Mechanical Engineering and Applied Mechanics (MEAM), Kevin Turner, Professor and Chair of MEAM with a secondary appointment in Materials Science Engineering, and their Ph.D. students, David Levine, Gokulanand Iyer and Daelan Roosa, published a study in Science Robotics describing a new, fracture-mechanics-based model of electroadhesive clutches, a mechanical structure that can control the stiffness of soft robotic materials.

Using this new model, the team was able to realize a clutch 63 times stronger than current electroadhesive clutches. The model not only increased force capacity of a clutch used in their soft robots, it also decreased the voltage required to power the clutch, making soft robots stronger and safer.

Current soft robotic hands can hold small objects, such as an apple for example. Being soft, the robotic hand can delicately grasp objects of various shapes, understand the energy required to lift them, and become stiff or tense enough to pick an object up, a task similar to how we grasp and hold things in our own hands. An electroadhesive clutch is a thin device that enhances the change of stiffness in the materials which allows the robot to perform this task. The clutch, similar to a clutch in a car, is the mechanical connection between moving objects in the system. In the case of electroadhesive clutches, two electrodes coated with a dielectric material become attracted to each other when voltage is applied. The attraction between the electrodes creates a friction force at the interface that keeps the two plates from slipping past each other. The electrodes are attached to the flexible material of the robotic hand. By turning the clutch on with an electrical voltage, the electrodes stick to each other, and the robotic hand holds more weight than it could previously. Turning the clutch off allows the plates to slide past each other and the hand to relax, so the object can be released.

Traditional models of clutches are based on a simple assumption of Coulombic friction between two parallel plates, where friction keeps the two plates of the clutch from sliding past each other. However, this model does not capture how mechanical stress is nonuniformly distributed in the system, and therefore, does not predict clutch force capacity well. It is also not robust enough to be used to develop stronger clutches without using high voltages, expensive materials, or intensive manufacturing processes. A robotic hand with a clutch created using the friction model may be able to pick up an entire bag of apples, but will require high voltages which make it unsafe for human interaction.

“Our approach tackles the force capacity of clutches at the model level,” says Pikul. “And our model, the fracture-mechanics-based model, is unique. Instead of creating parallel plate clutches, we based our design on lap joints and examined where fractures might occur in these joints. The friction model assumes that the stress on the system is uniform, which is not realistic. In reality, stress is concentrated at various points, and our model helps us understand where those points are. The resulting clutch is both stronger and safer as it requires only a third of the voltage compared to traditional clutches.”

“The fracture mechanics framework and model in this work have been used for the design of bonded joints and structural components for decades,” says Turner. “What is new here is the application of this model to the design of electroadhesive clutches.”

The researchers’ improved clutch can now be easily integrated into existing devices.

“The fracture-mechanics-based model provides fundamental insight into the workings of an electroadhesive clutch, helping us understand them more than the friction model ever could,” says Pikul. “We can already use the model to improve current clutches just by making very slight changes to material geometry and thickness, and we can continue to push the limits and improve the design of future clutches with this new understanding.”

To demonstrate the strength of their clutch, the team attached it to a pneumatic finger. Without the researchers’ clutch, the finger was able to hold the weight of one apple while inflated into a curled position; with it, the finger could hold an entire bag of them.

In another demonstration, the clutch was able to increase the strength of an elbow joint to be able to support the weight of a mannequin arm at the low energy demand of 125 volts.

Future work that the team is excited to delve into includes using this new clutch model to develop wearable augmented and virtual-reality devices.

“Traditional clutches require about 300 volts, a level that can be unsafe for human interaction,” says Levine. “We want to continue to improve our clutches, making them smaller, lighter and less energetically costly to bring these products to the real world. Eventually, these clutches could be used in wearable gloves that simulate object manipulation in a VR environment.”

“Current technologies provide feedback through vibrations, but simulating physical contact with a virtual object is limited with today’s devices,” says Pikul. “Imagine having both the visual simulation and feeling of being in another environment. VR and AR could be used in training, remote working, or just simulating touch and movement for those who lack those experiences in the real world. This technology gets us closer to those possibilities.”

Improving human-robot interactions is one of the main goals of Pikul’s lab and the direct benefits that this research presents is fuel for their own research passions.

“We haven’t seen many soft robots in our world yet, and that is, in part, due to their lack of strength, but now we have one solution to that challenge,” says Levine. “This new way to design clutches might lead to applications of soft robots that we cannot imagine right now. I want to create robots that help people, make people feel good, and enhance the human experience, and this work is getting us closer to that goal. I’m really excited to see where we go next.”

• PAPER – A mechanics-based approach to realize high–force capacity electroadhesives for robots. David J. Levine, Gokulanand M. Iyer, R. Daelan Roosa, Kevin T. Turner, and James H. Pikul. Science Robotics, 7(72), eabo2179.

Teleoperation is one of the longest-standing application fields in robotics. While full autonomy is still work in progress, the possibility to remotely operate a robot has already opened scenarios where humans can act in risky environments without endangering their own safety, such as when defusing explosives or decommissioning nuclear waste. It also allows one to be present and act even at great distance: underwater, in space, or inside a patient miles away from the surgeon. These are all critical applications, where skilled and qualified operators control the robot after receiving specific training to learn to use the system safely.

Teleoperation for everyone?

The recent pandemic has yet made even more apparent the need for immersive telepresence and remote action also for non-expert users: not only could teleoperated robots take vitals or bring drugs to infectious patients, but we could assist our elderly living far away with chores like moving heavy stuff, or cooking, for example. Also, numerous physical jobs could be executed from home.

The recent ANA-Xprize finals have shown how far teleoperation can go (see this impressive video of the winning team), but in such situations both the perceptual and control load lie entirely on the operator. This can be quite taxing on a cognitive level: both perception and action are mediated, by cameras and robotic arms respectively, reducing the user’s situation awareness and natural eye-hand coordination. While robot sensing capabilities and actuators have undergone relevant technological progress, the interface with the user still lacks intuitive solutions facilitating the operator’s job (Rea & Seo, 2022).

Human and robot joining forces

Shared control has gained popularity in recent years, as an approach championing human-machine cooperation: low-level motor control is carried out by the robot, while the human is focused on high-level action planning. To achieve such a blend, the robotic system still needs a timely way to infer the operator intention, so as to consequently assist with the execution. Usually, motor intentions are inferred by tracking arm movements or motion control commands (if the robot is operated by means of a joystick), but especially during object manipulation the hand is tightly following information collected by the gaze. In the last decades, increasing evidence in eye-hand coordination studies has shown that gaze reliably anticipates the hand movement target (Hayhoe et al., 2012), providing an early cue about human intention.

Gaze and motion features to estimate intentions

In a contribution presented at IROS 2022 last month (Belardinelli et al., 2022), we introduced an intention estimation model that relies on both gaze and motion features. We collected pick-and-place sequences in a virtual environment, where participants could operate two robotic grippers to grasp objects on a cluttered table. Motion controllers were used to track arm motions and to grasp objects by button press. Eye movements were tracked by the eye-tracker embedded in the virtual reality headset.

Gaze features were computed by defining a Gaussian distribution centered at the gaze position and taking for each object the likelihood for it to be the target of visual attention, which was given by the cumulative distribution collected by the object bounding box. For the motion features, the hand pose and velocity were used to estimate the hand’s current trajectory which was compared to an estimated optimal trajectory to each object. The normalized similarity between the two trajectories defined the likelihood of each object to be the target of the current movement.

Figure 1: Gaze features (top) and motion features (bottom) used for intention estimation. In both videos the object highlighted in green is the most likely target of visual attention and of hand movement, respectively.

These features along with the binary grasping state were used to train two Gaussian Hidden Markov Models, one on pick and one on place sequences. For 12 different intentions (picking of 6 different objects and placing at 6 different locations) the general accuracy (F1 score) was above 80%, even for occluded objects. Importantly, for both actions already 0.5 seconds before the end of the movement a prediction with over 90% accuracy was available for at least 70% of the observations. This would allow for an assisting plan to be instantiated and executed by the robot.

We also conducted an ablation study to determine the contribution of different feature combinations. While the models with gaze, motion, and grasping features performed better in the cross validation, the improvement with respect to only gaze and grasping state was minimal. Even when checking obstacles nearby at first, in fact, the gaze was already on the target before the hand trajectory became sufficiently discriminative.

We also ascertained that our models could generalize from one hand to the other (when fed the corresponding hand motion features), hence the same models could be used to concurrently estimate each hand intention. By feeding each hand prediction to a simple rule-based framework, basic bimanual intentions could also be recognized. So, for example, reaching for an object with the left hand while the right hand is going to place the same object on the left hand is considered a bimanual handover.

Figure 2: Online intention estimation: the red frame denotes the current right-hand intention prediction, the green frame the left-hand prediction. Above the scene, the bimanual intention is shown in capital letters.

Such an intention estimation model could help an operator to execute such manipulations without focusing on selecting the parameters for the exact motor execution of the pick and place, something we don’t usually do consciously in natural eye-hand coordination, since we automated such cognitive processes. For example, once a grasping intention is estimated with enough confidence, the robot could autonomously select the best grasp and grasping position and execute the grasp, relieving the operator of carefully monitoring a grasp without tactile feedback and possibly with inaccurate depth estimation.

Further, even if in our setup motion features were not decisive for early intention prediction, they might play a larger role in more complex settings and when extending the spectrum of bimanual manipulations.

Combined with suitable shared control policies and feedback visualizations, such systems could also enable untrained operators to control robotic manipulators transparently and effectively for longer times, improving the general mental workload of remote operation.

References

Belardinelli, A., Kondapally, A. R., Ruiken, D., Tanneberg, D., & Watabe, T. (2022). Intention estimation from gaze and motion features for human-robot shared-control object manipulation. 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022.

Hayhoe, M. M., McKinney, T., Chajka, K., & Pelz, J. B. (2012). Predictive eye movements in natural vision. Experimental brain research, 217(1), 125-136.

Rea, D. J., & Seo, S. H. (2022). Still Not Solved: A Call for Renewed Focus on User-Centered Teleoperation Interfaces. Frontiers in Robotics and AI, 9.

The Utah Bionic Leg is a motorized prosthetic for lower-limb amputees developed by University of Utah mechanical engineering associate professor Tommaso Lenzi and his students in the HGN Lab.

Lenzi’s Utah Bionic Leg uses motors, processors, and advanced artificial intelligence that all work together to give amputees more power to walk, stand-up, sit-down, and ascend and descend stairs and ramps. The extra power from the prosthesis makes these activities easier and less stressful for amputees, who normally need to over-use their upper body and intact leg to compensate for the lack of assistance from their prescribed prosthetics. The Utah Bionic Leg will help people with amputations, particularly elderly individuals, to walk much longer and attain new levels of mobility.

“If you walk faster, it will walk faster for you and give you more energy. Or it adapts automatically to the height of the steps in a staircase. Or it can help you cross over obstacles,” Lenzi says.

The Utah Bionic Leg uses custom-designed force and torque sensors as well as accelerometers and gyroscopes to help determine the leg’s position in space. Those sensors are connected to a computer processor that translates the sensor inputs into movements of the prosthetic joints. Based on that real-time data, the leg provides power to the motors in the joints to assist in walking, standing up, walking up and down stairs, or maneuvering around obstacles. The leg’s “smart transmission system” connects the electrical motors to the robotic joints of the prosthetic. This optimized system automatically adapts the joint behaviors for each activity, like shifting gears on a bike.

Finally, in addition to the robotic knee joint and robotic ankle joint, the Utah Bionic Leg has a robotic toe joint to provide more stability and comfort while walking. The sensors, processors, motors, transmission system, and robotic joints enable users to control the prosthetic intuitively and continuously, as if it was an intact biological leg.

Details of the leg’s newest technologies are described in a paper published in the journal. The paper was authored by University of Utah mechanical engineering graduate students Minh Tran, Lukas Grabert, Sarah Hood and Lenzi. You can read the paper here.

Lenzi and the university recently forged a new partnership with the worldwide leader in the prosthetics industry, Ottobock, to license the technology behind the Utah Bionic Leg and bring it to individuals with lower-limb amputations.

• PAPER – A lightweight robotic leg prosthesis replicating the biomechanics of the knee, ankle, and toe joint. Minh Tran†, Lukas Gabert, Sarah Hood, and Tommaso Lenzi. Science Robotics, 7(72), eabo3996.

This article was originally published here.

A newly published paper quantifies the extent to which automation has contributed to income inequality in the U.S., simply by replacing workers with technology — whether self-checkout machines, call-center systems, assembly-line technology, or other devices. Image: Jose-Luis Olivares, MIT

By Peter Dizikes

When you use self-checkout machines in supermarkets and drugstores, you are probably not — with all due respect — doing a better job of bagging your purchases than checkout clerks once did. Automation just makes bagging less expensive for large retail chains.

“If you introduce self-checkout kiosks, it’s not going to change productivity all that much,” says MIT economist Daron Acemoglu. However, in terms of lost wages for employees, he adds, “It’s going to have fairly large distributional effects, especially for low-skill service workers. It’s a labor-shifting device, rather than a productivity-increasing device.”

A newly published study co-authored by Acemoglu quantifies the extent to which automation has contributed to income inequality in the U.S., simply by replacing workers with technology — whether self-checkout machines, call-center systems, assembly-line technology, or other devices. Over the last four decades, the income gap between more- and less-educated workers has grown significantly; the study finds that automation accounts for more than half of that increase.

“This single one variable … explains 50 to 70 percent of the changes or variation between group inequality from 1980 to about 2016,” Acemoglu says.

The paper, “Tasks, Automation, and the Rise in U.S. Wage Inequality,” is being published in Econometrica. The authors are Acemoglu, who is an Institute Professor at MIT, and Pascual Restrepo PhD ’16, an assistant professor of economics at Boston University.

So much “so-so automation”

Since 1980 in the U.S., inflation-adjusted incomes of those with college and postgraduate degrees have risen substantially, while inflation-adjusted earnings of men without high school degrees has dropped by 15 percent.

How much of this change is due to automation? Growing income inequality could also stem from, among other things, the declining prevalence of labor unions, market concentration begetting a lack of competition for labor, or other types of technological change.

To conduct the study, Acemoglu and Restrepo used U.S. Bureau of Economic Analysis statistics on the extent to which human labor was used in 49 industries from 1987 to 2016, as well as data on machinery and software adopted in that time. The scholars also used data they had previously compiled about the adoption of robots in the U.S. from 1993 to 2014. In previous studies, Acemoglu and Restrepo have found that robots have by themselves replaced a substantial number of workers in the U.S., helped some firms dominate their industries, and contributed to inequality.

At the same time, the scholars used U.S. Census Bureau metrics, including its American Community Survey data, to track worker outcomes during this time for roughly 500 demographic subgroups, broken out by gender, education, age, race and ethnicity, and immigration status, while looking at employment, inflation-adjusted hourly wages, and more, from 1980 to 2016. By examining the links between changes in business practices alongside changes in labor market outcomes, the study can estimate what impact automation has had on workers.

Ultimately, Acemoglu and Restrepo conclude that the effects have been profound. Since 1980, for instance, they estimate that automation has reduced the wages of men without a high school degree by 8.8 percent and women without a high school degree by 2.3 percent, adjusted for inflation. 

A central conceptual point, Acemoglu says, is that automation should be regarded differently from other forms of innovation, with its own distinct effects in workplaces, and not just lumped in as part of a broader trend toward the implementation of technology in everyday life generally.

Consider again those self-checkout kiosks. Acemoglu calls these types of tools “so-so technology,” or “so-so automation,” because of the tradeoffs they contain: Such innovations are good for the corporate bottom line, bad for service-industry employees, and not hugely important in terms of overall productivity gains, the real marker of an innovation that may improve our overall quality of life.

“Technological change that creates or increases industry productivity, or productivity of one type of labor, creates [those] large productivity gains but does not have huge distributional effects,” Acemoglu says. “In contrast, automation creates very large distributional effects and may not have big productivity effects.”

A new perspective on the big picture

The results occupy a distinctive place in the literature on automation and jobs. Some popular accounts of technology have forecast a near-total wipeout of jobs in the future. Alternately, many scholars have developed a more nuanced picture, in which technology disproportionately benefits highly educated workers but also produces significant complementarities between high-tech tools and labor.

The current study differs at least by degree with this latter picture, presenting a more stark outlook in which automation reduces earnings power for workers and potentially reduces the extent to which policy solutions — more bargaining power for workers, less market concentration — could mitigate the detrimental effects of automation upon wages.

“These are controversial findings in the sense that they imply a much bigger effect for automation than anyone else has thought, and they also imply less explanatory power for other [factors],” Acemoglu says.

Still, he adds, in the effort to identify drivers of income inequality, the study “does not obviate other nontechnological theories completely. Moreover, the pace of automation is often influenced by various institutional factors, including labor’s bargaining power.”

Labor economists say the study is an important addition to the literature on automation, work, and inequality, and should be reckoned with in future discussions of these issues.

“Acemoglu and Restrepo’s paper proposes an elegant new theoretical framework for understanding the potentially complex effects of technical change on the aggregate structure of wages,” says Patrick Kline, a professor of economics at the University of California, Berkeley. “Their empirical finding that automation has been the dominant factor driving U.S. wage dispersion since 1980 is intriguing and seems certain to reignite debate over the relative roles of technical change and labor market institutions in generating wage inequality.”

For their part, in the paper Acemoglu and Restrepo identify multiple directions for future research. That includes investigating the reaction over time by both business and labor to the increase in automation; the quantitative effects of technologies that do create jobs; and the industry competition between firms that quickly adopted automation and those that did not.

The research was supported in part by Google, the Hewlett Foundation, Microsoft, the National Science Foundation, Schmidt Sciences, the Sloan Foundation, and the Smith Richardson Foundation.

• PAPER – Tasks, Automation, and the Rise in U.S. Wage Inequality. Daron Acemoglu, and Pascual Restrepo. Econometrica, 90: 1973-2016. https://doi.org/10.3982/ECTA19815

Researchers at MIT have made significant steps toward creating robots that could practically and economically assemble nearly anything, including things much larger than themselves, from vehicles to buildings to larger robots. The new system involves large, usable structures built from an array of tiny identical subunits called voxels (the volumetric equivalent of a 2-D pixel). Courtesy of the researchers.

By David L. Chandler

Researchers at MIT have made significant steps toward creating robots that could practically and economically assemble nearly anything, including things much larger than themselves, from vehicles to buildings to larger robots.

The new work, from MIT’s Center for Bits and Atoms (CBA), builds on years of research, including recent studies demonstrating that objects such as a deformable airplane wing and a functional racing car could be assembled from tiny identical lightweight pieces — and that robotic devices could be built to carry out some of this assembly work. Now, the team has shown that both the assembler bots and the components of the structure being built can all be made of the same subunits, and the robots can move independently in large numbers to accomplish large-scale assemblies quickly.

The new work is reported in the journal Nature Communications Engineering, in a paper by CBA doctoral student Amira Abdel-Rahman, Professor and CBA Director Neil Gershenfeld, and three others.

A fully autonomous self-replicating robot assembly system capable of both assembling larger structures, including larger robots, and planning the best construction sequence is still years away, Gershenfeld says. But the new work makes important strides toward that goal, including working out the complex tasks of when to build more robots and how big to make them, as well as how to organize swarms of bots of different sizes to build a structure efficiently without crashing into each other.

As in previous experiments, the new system involves large, usable structures built from an array of tiny identical subunits called voxels (the volumetric equivalent of a 2-D pixel). But while earlier voxels were purely mechanical structural pieces, the team has now developed complex voxels that each can carry both power and data from one unit to the next. This could enable the building of structures that can not only bear loads but also carry out work, such as lifting, moving and manipulating materials — including the voxels themselves.

“When we’re building these structures, you have to build in intelligence,” Gershenfeld says. While earlier versions of assembler bots were connected by bundles of wires to their power source and control systems, “what emerged was the idea of structural electronics — of making voxels that transmit power and data as well as force.” Looking at the new system in operation, he points out, “There’s no wires. There’s just the structure.”

The robots themselves consist of a string of several voxels joined end-to-end. These can grab another voxel using attachment points on one end, then move inchworm-like to the desired position, where the voxel can be attached to the growing structure and released there.

Gershenfeld explains that while the earlier system demonstrated by members of his group could in principle build arbitrarily large structures, as the size of those structures reached a certain point in relation to the size of the assembler robot, the process would become increasingly inefficient because of the ever-longer paths each bot would have to travel to bring each piece to its destination. At that point, with the new system, the bots could decide it was time to build a larger version of themselves that could reach longer distances and reduce the travel time. An even bigger structure might require yet another such step, with the new larger robots creating yet larger ones, while parts of a structure that include lots of fine detail may require more of the smallest robots.

Credit: Amira Abdel-Rahman/MIT Center for Bits and Atoms

As these robotic devices work on assembling something, Abdel-Rahman says, they face choices at every step along the way: “It could build a structure, or it could build another robot of the same size, or it could build a bigger robot.” Part of the work the researchers have been focusing on is creating the algorithms for such decision-making.

“For example, if you want to build a cone or a half-sphere,” she says, “how do you start the path planning, and how do you divide this shape” into different areas that different bots can work on? The software they developed allows someone to input a shape and get an output that shows where to place the first block, and each one after that, based on the distances that need to be traversed.

There are thousands of papers published on route-planning for robots, Gershenfeld says. “But the step after that, of the robot having to make the decision to build another robot or a different kind of robot — that’s new. There’s really nothing prior on that.”

While the experimental system can carry out the assembly and includes the power and data links, in the current versions the connectors between the tiny subunits are not strong enough to bear the necessary loads. The team, including graduate student Miana Smith, is now focusing on developing stronger connectors. “These robots can walk and can place parts,” Gershenfeld says, “but we are almost — but not quite — at the point where one of these robots makes another one and it walks away. And that’s down to fine-tuning of things, like the force of actuators and the strength of joints. … But it’s far enough along that these are the parts that will lead to it.”

Ultimately, such systems might be used to construct a wide variety of large, high-value structures. For example, currently the way airplanes are built involves huge factories with gantries much larger than the components they build, and then “when you make a jumbo jet, you need jumbo jets to carry the parts of the jumbo jet to make it,” Gershenfeld says. With a system like this built up from tiny components assembled by tiny robots, “The final assembly of the airplane is the only assembly.”

Similarly, in producing a new car, “you can spend a year on tooling” before the first car gets actually built, he says. The new system would bypass that whole process. Such potential efficiencies are why Gershenfeld and his students have been working closely with car companies, aviation companies, and NASA. But even the relatively low-tech building construction industry could potentially also benefit.

While there has been increasing interest in 3-D-printed houses, today those require printing machinery as large or larger than the house being built. Again, the potential for such structures to instead be assembled by swarms of tiny robots could provide benefits. And the Defense Advanced Research Projects Agency is also interested in the work for the possibility of building structures for coastal protection against erosion and sea level rise.

The new study shows that both the assembler bots and the components of the structure being built can all be made of the same subunits, and the robots can move independently in large numbers to accomplish large-scale assemblies quickly. Courtesy of the researchers.

Aaron Becker, an associate professor of electrical and computer engineering at the University of Houston, who was not associated with this research, calls this paper “a home run — [offering] an innovative hardware system, a new way to think about scaling a swarm, and rigorous algorithms.”

Becker adds: “This paper examines a critical area of reconfigurable systems: how to quickly scale up a robotic workforce and use it to efficiently assemble materials into a desired structure. … This is the first work I’ve seen that attacks the problem from a radically new perspective — using a raw set of robot parts to build a suite of robots whose sizes are optimized to build the desired structure (and other robots) as fast as possible.”

The research team also included MIT-CBA student Benjamin Jenett and Christopher Cameron, who is now at the U.S. Army Research Laboratory. The work was supported by NASA, the U.S. Army Research Laboratory, and CBA consortia funding.

• PAPER – Self-replicating hierarchical modular robotic swarms. Amira Abdel-Rahman, Christopher Cameron, Benjamin Jenett, Miana Smith, and Neil Gershenfeld. Commun Eng 1, 35 (2022). https://doi.org/10.1038/s44172-022-00034-3

Magnetic bacteria (grey) can squeeze through narrow intercellular spaces to cross the blood vessel wall and infiltrate tumours. (Visualisations: Yimo Yan / ETH Zurich)

By Fabio Bergamin

Scientists around the world are researching how anti-cancer drugs can most efficiently reach the tumours they target. One possibility is to use modified bacteria as “ferries” to carry the drugs through the bloodstream to the tumours. Researchers at ETH Zurich have now succeeded in controlling certain bacteria so that they can effectively cross the blood vessel wall and infiltrate tumour tissue.

Led by Simone Schürle, Professor of Responsive Biomedical Systems, the ETH Zurich researchers chose to work with bacteria that are naturally magnetic due to iron oxide particles they contain. These bacteria of the genus Magnetospirillum respond to magnetic fields and can be controlled by magnets from outside the body; for more on this, see an earlier article in ETH News.

Exploiting temporary gaps

In cell cultures and in mice, Schürle and her team have now shown that a rotating magnetic field applied at the tumour improves the bacteria’s ability to cross the vascular wall near the cancerous growth. At the vascular wall, the rotating magnetic field propels the bacteria forward in a circular motion.

To better understand the mechanism to cross the vessel wall works, a detailed look is necessary: The blood vessel wall consists of a layer of cells and serves as a barrier between the bloodstream and the tumour tissue, which is permeated by many small blood vessels. Narrow spaces between these cells allow certain molecules from the to pass through the vessel wall. How large these intercellular spaces are is regulated by the cells of the vessel wall, and they can be temporarily wide enough to allow even bacteria to pass through the vessel wall.

Strong propulsion and high probability

With the help of experiments and computer simulations, the ETH Zurich researchers were able to show that propelling the bacteria using a rotating magnetic field is effective for three reasons. First, propulsion via a rotating magnetic field is ten times more powerful than propulsion via a static magnetic field. The latter merely sets the direction and the bacteria have to move under their own power.

The second and most critical reason is that bacteria driven by the rotating magnetic field are constantly in motion, travelling along the vascular wall. This makes them more likely to encounter the gaps that briefly open between vessel wall cells compared to other propulsion types, in which the bacteria’s motion is less explorative. And third, unlike other methods, the bacteria do not need to be tracked via imaging. Once the magnetic field is positioned over the tumour, it does not need to be readjusted.

“Cargo” accumulates in tumour tissue

“We make use of the bacteria’s natural and autonomous locomotion as well,” Schürle explains. “Once the bacteria have passed through the blood vessel wall and are in the tumour, they can independently migrate deep into its interior.” For this reason, the scientists use the propulsion via the external magnetic field for just one hour – long enough for the bacteria to efficiently pass through the vascular wall and reach the tumour.

Such bacteria could carry anti-cancer drugs in the future. In their cell culture studies, the ETH Zurich researchers simulated this application by attaching liposomes (nanospheres of fat-like substances) to the bacteria. They tagged these liposomes with a fluorescent dye, which allowed them to demonstrate in the Petri dish that the bacteria had indeed delivered their “cargo” inside the cancerous tissue, where it accumulated. In a future medical application, the liposomes would be filled with a drug.

Bacterial cancer therapy

Using bacteria as ferries for drugs is one of two ways that bacteria can help in the fight against cancer. The other approach is over a hundred years old and currently experiencing a revival: using the natural propensity of certain species of bacteria to damage tumour cells. This may involve several mechanisms. In any case, it is known that the bacteria stimulate certain cells of the immune system, which then eliminate the tumour.

“We think that we can increase the efficacy of bacterial cancer therapy by using a engineering approach.”

– Simone Schürle

Multiple research projects are currently investigating the efficacy of E. coli bacteria against tumours. Today, it is possible to modify bacteria using synthetic biology to optimise their therapeutic effect, reduce side effects and make them safer.

Making non-magnetic bacteria magnetic

Yet to use the inherent properties of bacteria in cancer therapy, the question of how these bacteria can reach the tumour efficiently still remains. While it is possible to inject the bacteria directly into tumours near the surface of the body, this is not possible for tumours deep inside the body. That is where Professor Schürle’s microrobotic control comes in. “We believe we can use our engineering approach to increase the efficacy of bacterial cancer therapy,” she says.

E. coli used in the cancer studies is not magnetic and thus cannot be propelled and controlled by a magnetic field. In general, magnetic responsiveness is a very rare phenomenon among bacteria. Magnetospirillum is one of the few genera of bacteria that have this property.

Schürle therefore wants to make E. coli bacteria magnetic as well. This could one day make it possible to use a magnetic field to control clinically used therapeutic bacteria that have no natural magnetism.

• PAPER – Magnetic torque–driven living microrobots for increased tumor infiltration. Gwisai T, Mirkhani N, Christiansen MG, Nguyen TT, Ling V, and Schuerle S. Science Robotics, vol.7(71)