By the University of Glasgow

Giving robot companions animal-like or anime-style faces and voices could maximise their emotional appeal to users, according to new research which could help guide the growing field of social robotics.

Social robots can provide companionship and comfort to their owners by mimicking human or petlike behaviours. Although the robots are a fast-growing sector of the tech industry, surprisingly little academic research has explored how their designs can be optimised to make the robots as expressive and emotionally engaging as possible.

In a new paper set to be presented at a conference this week, researchers from the University of Glasgow show how they have been working to close that research gap.

With the help of a group of volunteers, they used augmented reality technology to test dozens of different combinations of faces and voices overlaid on the body of an existing social robot called a Qoobo, which has a soft, cushion-like body and a swinging tail.

Dr Shaun Macdonald and Josh Yip of the University of Glasgow’s School of Computing Science used this rapid prototyping approach to explore how users felt about the virtual robots, and how effectively they were able to ‘read’ the robot’s emotional states to make them feel more like real creatures.

Dr Macdonald said: “Social robots can provide valuable experiences for a wide range of people. Older people have reported that they can help mitigate feelings of loneliness or social isolation, for example, and people living in spaces where they can’t keep a real pet have kindled some of that sense of companionship with what we call zoomorphic robots, which mimic animals’ appearances and behaviour.

“However, the aesthetic design of zoomorphic social robots so far has largely been the result of educated guesswork, with very little research available on what makes them appealing to users.

“What we’ve done here for the first time here is take a more systematic approach to designing the appearance of a zoomorphic robot by harnessing augmented reality to enable rapid, side-by-side design tests. In the future, these insights could help create a new generation of user-customisable robots which can tailor their appearances to better suit their owners’ preferences.”

The research builds on Dr Macdonald’s previous research in advancing the design of social robots. His work includes the development of a software system called Augmenting Zoomorphic Robotics with Affect, or AZRA, which can overlay augmented-reality sounds and animations over the bodies of real-world social robots to prototype new appearances and functionalities.

In this research, the team began by creating a ‘mood board’ of appealing images of animals and robots drawn from real life and from popular culture to help them design a series of prototype faces for virtual robots. They ended up with five prototype faces comprised of animal-like, robot-like, emoji-style, anime-inspired designs, with another comprised simply of eyes.

They chose five distinct vocalisations to enable the prototypes to express themselves to the volunteers: recordings of cat noises, human-like vocal sounds, abstract electronic noises, music and samples of ‘animalese’, the nonsense sounds used by the characters in the popular Nintendo game series Animal Crossing.

Then, 24 volunteers were given AR headsets and asked to rank different combinations of faces and voices overlaid on Qoobo for how clearly they expressed an emotion and how much empathy they felt towards the prototype robot, as well as how appropriate the combination of factors felt to them.

The results revealed that combinations of animal-like faces and vocalisations matched Qoobo’s animal-like body were well-liked by the volunteers, who found them more engaging and immersive. While that result might be unsurprising, the team also found that anime-like faces were just as highly-ranked despite not being based on real life.

Josh Yip said: “Anime-like faces are more expressive than animal faces could ever really be, giving the robot abilities to express themselves beyond those of a normal pet but in a way that didn’t feel uncanny. It still felt like it could be a pet-like creature to our study participants.

“We also found that faces consistently outperformed sounds at conveying emotion and creating connections. Interestingly, the study participants found that the human sounds were easier to read emotionally but were liked less than animal sounds, even though they found them harder to clearly identify with specific emotions.”

The team were surprised to find that, although the study participants insisted that they felt that being able to correctly read the robot’s emotions was very important to them, the data showed that readability had little influence on the designs they preferred. Along similar lines, ‘animalese’ sounds ranked highly with users despite being unintelligible, with some users feeling it suggested the robot had a greater sense of intelligence than other sounds.

The study’s participants reactions to the prototypes were also affected by whether they owned a real animal pet themselves. Pet owners were more likely to treat the animal-like prototypes as real pets, but to be more likely to treat robot-like faces as machines.

Dr Macdonald added: “Our research suggests that although there is no single face or voice that has universal appeal, the combination of the two can have a big effect on how well users feel that robots are able to communicate emotions.

“We hope our findings will help inform the design of future generations of zoomorphic social robot technology, but they could also extend the lifespan of existing devices by enabling users to overlay new faces, voices and functionality on them using augmented reality.”

The team’s paper, titled ‘Anime, Animal or Animal Crossing? Comparing Aesthetic Styles for Zoomorphic Robot Faces and Sounds using Augmented Reality’, will be presented at the ACM International Conference on Mobile Human-Computer Interaction conference in Swansea, Wales on Thursday 3 September.

The project was supported by funding from the University of Glasgow’s EPSRC Impact Acceleration Account and the Engineering and Physical Sciences Research Council (EPSRC).

This article is republished from the University of Glasgow under a Creative Commons license. Read the original article here.

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).

The 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2026) will be held from 27 September – 1 October in Pittsburgh, USA. The programme includes plenary and keynote talks, workshops, tutorials, forums, competitions, and a debate.

Keynote talks

The keynote talks will be held from 28 – 30 September:

• Elena De Momi – Open Questions on the Road to Autonomous Surgery
• Naomi Fitter – Socially-Informed Context Awareness for Robots in Human-Populated Settings
• Greg Chirikjian – From Hyper-Redundancy to Robot Imagination and Affordance Learning
• Genta Suzuki – Fujitsu Kozuchi Physical OS: A Research Strategy for Scalable Physical AI
• Gaurav Bansal – Title to be confirmed
• Al Rizzi – Developing Advanced Robot Capability at the RAI Institute
• Yekai Sun – Tokens of the Physical World: Accelerating Intelligence Emergence via Multimodal Embodied Data
• He Wang – Towards the AlphaGo and ChatGPT Moments of Embodied AI
• Jen Jen Chung – All data are equal, but some data are more equal than others
• Yigit Menguc – Soft Robotics, Tensegrity, and the People Who Love Them
• Robin Murphy – Rethinking Disaster Robotics: The Open Questions That Matter Now
• Gina Olson – Embracing Design Complexity: A Biomimetic Approach to (Soft?) Snake Robots
• Tim Bretl – How you can be part of a revolution in composite manufacturing
• Lael Odhner – Big bets in manipulation systems: where is the frontier today?
• David Held – Robots That Learn After Deployment: Designing Policies for Fast Adaptation
• Ankur Mehta – Robotics problems that robotics can’t solve
• Brittany Duncan – Patterns from the Field: how contextual lessons can inform robot design
• Ralph Hollis – Lorentz Levitation Technology for Fine Motion Robotics, Haptic Interaction, and Vibration Isolation
• Genki Yamauchi – From Big Sites to Every Site: Extending Construction Robotics to the Work Done by Hand
• Nancy Pollard – Title to be confirmed

Forums

Community discussion sessions open to all attendees, these forums will be held on Monday 28 and Tuesday 29 September.

• Monday 28 (afternoon): US Government Funding
• Tuesday 29 (morning): Robotics Education: Experiential Learning
• Tuesday 29 (afternoon): Is Robotics Underinvesting in Its Own Foundations?

Plenary panels

Three plenary panels will be held:

• Monday 28: Start-ups and entrepreneurship
• Tuesday 29: Space
• Wednesday 30: Should robots be generalists or specialists?

Workshops and tutorials

There are 44 workshops and tutorials planned for Sunday 27 September, with a further 42 scheduled for Thursday 1 October. You can find out more about these here.

Introduction to science communication

Join Sabine Hauert (University of Bristol and Robohub), Ella Scallan (Robohub), and Evan Ackerman (IEEE Spectrum) for a science communication crash course. This will take place at 11:00 on Tuesday 29 September. Find out more here.

Find out more

• Conference webpage
• Technical tours
• Competitions

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 Ashwin R. Vasavada

Lights blink on as I enter the Rover Operations Center at NASA’s Jet Propulsion Laboratory in Pasadena, California, at 7:30 a.m. I’m the first to arrive, even though I already feel late.

Sometime during the night on Earth, the Curiosity rover finished its day exploring on Mars and beamed its latest collection of images and measurements to a Mars orbiter zooming by overhead. The orbiter relayed the data to Earth, where it now waits on servers here and at partner institutions around the globe.

Once our rover operations shift kicks off at 8:15 a.m., a few dozen engineers and scientists will have just three hours to check the health of the rover, analyze the new science data and agree on the next set of rover activities. Then we’ll spend another four hours turning those plans into rover commands for tomorrow, making sure they are safe and fit within the rover’s available time and energy. Not long after that, the Sun will rise on Mars, and Curiosity will look toward Earth, expecting its next instructions.

As Curiosity’s project scientist, it’s my job to ensure that what emerges from this rush is a set of measurements that advance the mission’s science objectives and keep it on track to achieve what our team promised NASA and, ultimately, the public. It’s a seemingly overwhelming task given everything that must happen in the next seven hours. But after repeating it over a thousand times since Curiosity landed in 2012, our team has gotten pretty good at it.

NASA Curiosity project scientist Ashwin Vasavada guides this tour of the rover’s view of the Martian surface.

Reading the rocks

NASA created Curiosity to search for evidence of ancient habitable environments, such as those with liquid water and the chemicals, nutrients and energy sources required for life. The investigation called for a long-lived, mobile spacecraft that would allow scientists on Earth to virtually explore a local area on Mars, select and acquire rock samples, and analyze them in onboard laboratories.

JPL responded with the car-size, drill-equipped Curiosity rover. NASA added a suite of scientific instruments and a science team from the United States and around the world.

NASA sent Curiosity to Mars’ Gale Crater to climb Aeolis Mons, a mountain whose 3 miles (5 kilometers) of sedimentary rock layers hold a record of environmental conditions from about 3.5 billion years ago. The evidence suggests that back then a thicker, ancient atmosphere sheltered flowing streams and sparkling lakes.

Curiosity has climbed through a vertical half-mile (1 kilometer) of rock layers so far, finding clay-rich, mudstone layers that give way at higher elevations to younger sandstones full of salty minerals. Our team determined that lakes persisted for millions of years before the climate became arid and sand dunes overtook the lakes, although groundwater occasionally breached the surface to produce streams that wove among the dunes.

Samples Curiosity drilled from the lake sediments contain small organic – meaning carbon-based – molecules, the raw materials for potential life. The association of wet environments, organic molecules and a mix of chemicals similar to those that microbes harness for energy on Earth allowed our team to conclude that conditions in Gale were once capable of supporting life. Determining conclusively whether life actually took hold will require bringing such rocks back to laboratories on Earth.

Preparing a plan

As my colleagues arrive in the building and online, I’m lost in the latest images. Curiosity is well into the higher and drier strata of Aeolis Mons, yet the rocks have salts, scours and other signs of ancient water. These images suggest that perhaps even Mars’ dry period provided habitats for life.

During our previous shift, I asked the engineers who plan the rover’s route to take it into an area that appears exceptionally smooth and flat on images taken from orbit. After years of rough offroad driving, they were excited to open up the throttle and let the rover drive 120 feet (37 meters) across this Martian “parking lot” – a pretty long distance for Curiosity.

Polygon fractures with honeycomblike textures, discovered by NASA’s Curiosity Mars rover. NASA/JPL-Caltech/MSSS

What I’m looking at now is taking my breath away. Instead of smooth slabs, every surface around the rover is textured by small polygons, each a few inches across, as if someone covered Mars in honeycomb wallpaper. Our geologists suspect that the polygons outline fractures that formed through drying, like cracks in mud do, or through thermal cycles, or compression.

My computer screen shows the area in front of Curiosity covered by dots, each one representing a candidate target for the rover to observe with its cameras, its laser spectrometer, or the sensors on its arm. There are way too many dots, which is not a bad problem to have. A science team member trained to moderate the day’s discussion begins to narrow the list. Distinguished faculty, postdocs and students from around the world take turns advocating for their target.

The debate centers on what set of images and chemical measurements will best help distinguish how the polygons formed. I’m the tiebreaker if our team can’t quickly reach consensus, but that’s rarely needed. It’s Friday, so the three days of activities we plan for the rover today will take place over the weekend, and the results will arrive for our next shift on Monday. Curiosity works weekends.

Engineering the uplink

Later on, I walk into the adjacent room where a robotics engineer visualizes in 3D how Curiosity’s five-jointed arm will reach the winning rock targets. I sit down next to another engineer who is simulating the rover’s next drive. We agree that the terrain ahead is rough: no more parking lot, all curbs. We discuss how she might steer the rover around the rock obstacles to reach the next science waypoint.

By noon, the science team’s role in operations is done, but my colleagues linger online to discuss the latest data in more detail. Meanwhile, the rover and instrument operators begin converting the day’s scientific requests into hundreds of commands that must be sent to NASA’s Deep Space Network in a few hours. There’s still a lot to do, but the team is in the homestretch for the day.

On Mars, the horizon is starting to brighten. Before too long, Curiosity will be expecting to hear from us.

This research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract (80NM0018D0004) with the National Aeronautics and Space Administration.

By Emmet Cole

In our new monthly series, we showcase a selection of robotics stories from society, business, research, and culture, as we track the field’s ongoing journey from specialized industrial machines to an increasingly visible social phenomenon.

Reshoring reimagined?

In early August, The Information published an article that claimed U.S.-based founders, investors and startups, are travelling to an electronics market in Shenzhen to buy servomotors, sensors, controllers and gimbals. The report also claimed that Unitree and Zhiyuan robots are being bought intact, then stripped down into parts and carried back to the USA in suitcases.

By late August, in a separate development, reports emerged that a US Department of Energy lab is investigating whether Chinese LiDAR sensors pose a security risk, if widely used on vehicles in the United States. According to TechCrunch, the research is being funded by one or more companies in the electric and autonomous vehicle industries.

Together, the stories highlight the ongoing tension between US security concerns and its robotics industry’s continuing dependence on Chinese hardware.

Motorless shape shifters

Don’t cancel any Shenzhen travel plans just yet, but engineers at Princeton have unveiled a motorless, origami-inspired robot that can roll, crawl, and change shape through a combination of magnetic control and multistable geometry.

Project lead Glaucio Paulino told Princeton Engineering:

Geometry is the real actuator here. Instead of relying on complicated mechanisms, we use mathematical principles to encode multiple stable configurations directly into the structure. That opens a new pathway for designing lightweight, adaptable systems.

Managing expectations

Robot deployments can fail for many reasons, making expectation management a key part of successful robotics projects. This might mean explaining that although collaborative robots can sometimes be deployed without fencing, that is not appropriate for every application, and a risk assessment is still required.

It might mean explaining that although robot-assisted radical prostatectomy typically reduces blood loss and shortens hospital stays compared with open surgery, longer-term outcomes depend heavily on the patient, the cancer and the surgeon.

In the humanoid space, where expectations are often formed by spectacle and driven by anthropomorphism and millions of years of evolution, the effects can be particularly acute.

A Drexel-led study involving 50 adult men that measured brain activity, hormones, self-reported attitudes and behaviour in order to assess trust in human-humanoid interaction, found that participants formed stronger initial connections with an expressive Pepper robot than with a stationary version that gave no nonverbal cues.

But when the expressive robot made errors, the participants’ brains responded as they would to a person breaking a social norm, rather than a machine making a technical error.

Counterintuitively:

[…] levels of “the bonding hormone,” oxytocin, which is typically present in higher amounts among friends, relatives and romantic partners, actually rose in participants as the engaging robot made mistakes. This led [researchers] to believe that the hormone functions as a warning signal, rather than a sign of connection, in human-robot interactions.

Read the full Science Robotics paper here.

Beyond automotive

For decades, the automotive sector has been the leading segment for robotics adoption. That’s still the case, but as the latest North American Q2 figures from the Association for Advancing Automation (A3) show, automotive OEM’s share is continuing to slide as robots diversify across other industry segments.

Notably, North American companies ordered 8,940 robots valued at US$622 million in the second quarter of 2026, a 4.3% increase in units ordered and a 21.3% increase in revenue, according to A3’s numbers.

Recycling robots

IEEE Spectrum profiled an automated recycling system developed by researchers at Germany’s Karlsruhe Institute of Technology. The system uses CAD data, physical observations and a predictive algorithm to predict defects in products and adapt its disassembly strategy while protecting valuable components.

All eyes on China

As August came to a close, the eyes of the robotics world (and the general public) turned to Beijing, China as the city played host to both the World Robot Conference (WRC) and the World Humanoid Robot Games.

The WRC, primarily a showcase of Chinese products, brought together ~300 companies exhibiting more than 2,000 products. These included robots for parcel sorting, manufacturing, surgery and domestic work.

Unitree supplied much of the spectacle with its remarkably fast “Superman” humanoid and an extraordinary stock-market debut. Its shares closed 460% above the IPO price, although by August 28 they had fallen almost 30% -another lesson in managing expectations. CEO Wang Xingxing also predicted that robotics is approaching a “ChatGPT moment”.

The World Humanoid Robot Games brought the theme of spectacle versus reality to another level. Robots kick-boxed, played table tennis, performed gymnastic routines, and played football. There were moments of hilarity…

…and some genuinely impressive technology demonstrations.

After watching several hours of livestreams from the games across several days, and looking at the print media coverage, I began to wonder whether events like these help people manage expectations around robots much.

The media focus on spectacle doesn’t help. Yes, a humanoid ran at Usain Bolt-like speeds and then into a wall. Yes, the kickboxing humanoids had an endearing Mr. Magoo-like quality. Technology demonstration aside, there isn’t much call for either in industrial applications.

Blink and you’d miss it -and much of the media coverage did- but the games also hosted multiple competitions centred around industrial manufacturing, logistics, household services and emergency response.

In the industrial “assembly and material supply” contest, for example, robots had to move containers onto shelves, identify and sort differently packaged components, and insert intake and exhaust valves into the correct openings in an engine cylinder head.

Other contests tested packaging and warehouse intake, while eight dexterous-hand challenges included screw fastening, unpacking boxes, connecting cables and picking up beans with tweezers. Some competitions required autonomous robots and others tested teleoperation capabilities.

These slower, less spectacular competitions tested the precision, perception and adaptability humanoids will need before they can perform useful work in factories and warehouses.

All that said, everybody needs a humanoid riding a giant quadruped.

By Ellen H. Rumley and Allison Okamura

A leader in tech innovation

Robots are a crowd-pleasing example of American innovation and global technological leadership. The latest press release of Boston Dynamics’ Atlas executing the ‘Ghost Rabona’ soccer kick highlights the increasingly natural movements of humanoids (1). At the other end of the spectrum, academic researchers are redefining the limits of miniaturization and building robots at the micrometer scale (2). Private and government funding propel foundational research forward, leading to breakthroughs in both hardware and software across U.S. labs and corporations (3, 4). Such progress reaffirms the U.S. as an intellectual and innovative powerhouse, a reputation that carries large geopolitical weight.

But retaining a lead solely in the R&D phase yields diminishing returns if the U.S. fails to manufacture robotic hardware at scale within its own borders. Critics fear this dynamic prevents the U.S. from reaping the economic benefits of the very technology it pioneers (5, 6). Ramping up domestic production will require a heavily coordinated effort between government, industry, and academics – a pivot from traditional U.S. technology strategy.

Free market and military capital

Historically, the U.S. has maintained a laissez-faire approach to technological innovation, allowing corporate and private investors driven by the free market to shape the course of advancement. This innovation-first strategy is largely credited with fueling the growth of global tech giants and the rise of Silicon Valley, which altogether represents roughly 12% of the national GDP (7). America’s leadership in artificial intelligence can be largely attributed to this free-market engine, which prioritizes high projected profit margins and rapid scalability of software (8).

When it comes to U.S. robotics hardware, the Department of Defense/War (DoD/DoW) has stepped in as one of the primary catalysts for innovation and investment. The Pentagon’s fiscal year 2026 budget requested an unprecedented $13.4 billion for developing autonomous systems; a major increase from years prior. The DoD/DoW is requesting an additional $53.6 billion for autonomous systems and drones for 2027 (9). Private capital has quickly mirrored federal priorities; in the first quarter of 2026 alone, defense technology venture capital investments reached a record $19.8 billion. This surge is tightly bound to contemporary geopolitical conflicts, such as the hybrid warfare seen in Ukraine, where the U.S. plays a key role in funding drone production and in recent times deploying humanoids (10, 11). The significant role that autonomous systems have played in conflict zones like the Strait of Hormuz has also signaled the high dependency of strategic military operations on robotics (12).

A bottleneck for domestic manufacturing

While investment in robotic innovation does provide certain economic benefits, it does not support U.S. competitiveness in critical industries – specifically in physical manufacturing. Hardware inherently requires higher, long-term capital expenditures and yields lower short-term financial returns compared to software. Consequently, there could be less incentive to compete against other global manufacturing leaders. Furthermore, while the DoD funds early-stage R&D and specialized military platforms, its mandate stops short of supporting commercial scaling efforts necessary to implement technologies for broader civilian and industrial applications.

As a result, while the U.S. is the third-largest consumer of industrial robotics in the world (with industrial adoption surging by 11% just last year in sectors like logistics, packaging and automotive) (13), it lacks the infrastructure to build what it consumes. While a few prominent U.S. companies do manufacture application-specific robots (some examples being Intuitive Surgical, Amazon Robotics, and Agility), the U.S. mostly relies on foreign imports for baseline industrial robot installations (14). Contrast this with Japan, which relies on foreign imports for a mere 2% of its domestic installations (15).

The systemic under-development in domestic robotic manufacturing leaves the U.S. vulnerable to hardware dependencies and supply chain disruptions. While industrial leaders like Japan and Germany are reliable trade partners, the U.S. government increasingly views hardware dependencies as a national security vulnerability. In recent years, robots have been reclassified as a critical dual-use technology, considered central to 21st century technological sovereignty (16). This reality makes robotics manufacturing a target for foreign adversaries, since foreign supply chains introduce operational chokepoints.

Currently, the U.S. government responds to these vulnerabilities through aggressive export controls and protective procurement policies, which are designed to slow down adversary agendas by blocking their access to the U.S. IP and markets. As one example, in 2021 the President issued the NSPM-33 (a national security memorandum directed to executive agencies) mandating federally funded research institutions to establish strict internal compliance frameworks (17). While designed to allow the continued recruitment of international talent, it enforces rigid vetting protocols on researchers potentially associated with Malign Foreign Talent Recruitment Programs (MFTRP) or of dual national affiliation (18). As another example, in early 2026 lawmakers proposed the bipartisan American Security Robotics Act, a bill designed to ban the federal procurement and operation of unmanned ground vehicles and humanoids manufactured by foreign adversaries, specifically targeting China (19).

Getting by without a government-led national robotics strategy

Rising concerns over robotics supply chain vulnerabilities strongly echo the semiconductor anxieties that originally led to the CHIPS and Science Act of 2022, which marked a historical pivot for the U.S. towards an active state-supported industrial policy (20). Today, there is a similar pressure mounting for the federal government to coordinate a sovereign robotics pipeline via a unified national roadmap.

The challenge facing the U.S. does not seem to be a lack of federal funding, but rather how the funding is distributed across multiple agencies with different mandates (21). The National Science Foundation funds basic academic research; Department of Energy runs national research laboratories and funds external research programs developing robotics relating to energy infrastructure; National Institute of Health supports and conducts medical robotics research; NASA funds and develops robotics for space applications; the ARM Institute supports domestic manufacturing initiatives while satisfying dual-use funding criteria through the DoD/DoW; and DARPA funds robotic innovations relevant to national security. (The latter organizes the well-known DARPA Grand Challenges, robotics competitions for fueling research bridging fundamental science and military applications.) Because federal investments from these programs and agencies are confined to isolated, mission-specific niches, the U.S. remains one of the few leading industrial nations without a centralized robotics strategy.

This structural fragmentation may soon change as political attention shifts towards treating robotics as a critical sector (22). In 2025, four state representatives re-launched the bipartisan Congressional Robotics Caucus as a platform for members of Congress to stay informed over the diverse issues surrounding robotics policy (23). In June 2026, legislators formally introduced the bipartisan National Commission on Robotics Act, which calls for an independent commission of 18 robotics experts tasked with providing evidence-based policy to accelerate domestic development (24). This legislative momentum is reinforced by a series of recent Executive Orders (mandates from the US President) demanding a high-level federal coordination including the Unleashing American Drone Dominance Order (25), the Golden Dome for America Order (26), and the Launching the Genesis Mission Order (27).

Robotics strategies from non-government actors

In the absence of a centralized federal initiative, stakeholders in industry, NGOs, and academia are self-organizing and publishing their own robotics frameworks recommendations for the United States.

In 2025, A3 (Association for Advancing Automation) released their Vision for a U.S. National Robotics and Automation Strategy (28), which argues that the future of U.S. leadership in artificial intelligence is inextricably linked with its global robotics leadership. A3 provides actionable policy items, such as the formation of a governmental robotics office for coordinating federal robotics initiatives, the establishment new standards for nascent robotic hardware and software, and the introduction of tax incentives for rapid robotics adoption and creation of worker training programs.

Complementing these industry goals, in 2025 the nonpartisan thinktank, Special Competitive Studies Project (SCSP), has released their Memos to the President – National Robotics Strategy (29). This document couples the call for mass industrial robotics adoption with recommendations for more aggressive trade restrictions on Chinese technologies. Following this publication, SCSP launched its National Security Commission on Robotics for Advanced Manufacturing for implementing this comprehensive strategy; members include industry and academic players including GM, Boston Dynamics, Nvidia, AMD, the University of Michigan and MIT’s Industrial Performance Center (30).

Academics have similarly spearheaded their own national strategies. Once every four years, academics led by Professor Henrik Christensen (UC San Diego) publish a roadmap for U.S. robotics (31). Past editions directly catalyzed the creation of the National Robotics Initiative (NRI), a multi-agency federal program running from 2012 to 2022 that supported foundational research on robotics, specifically promoting collaborative robots for working alongside humans. (Christensen also recently published Global Robotics Technology Roadmap 2025–2035, an independent positions paper synthesizing government and industry strategies across Europe, Asia, and the United States (32).)

All three stakeholders share a common motivation for a national robotics strategy: the nation’s population and workforce are declining; There is a surplus of jobs fulfilling the 3 Ds (dull, dirty, dangerous); Robotics should become a national priority once more. In addition to funding and incentives, authors also discuss the need to consider the impact of technology on the workforce – a topic of large structural tension, with fear of robots displacing jobs and particularly outcompeting blue-collar and immigrant workers (33, 34). Christensen et al. delve deeper into the psychology of the workforce, and recommend a framework which balances technological growth with worker empowerment. They propose that prioritizing human-centered collaborative robots and wearable devices could shift the robotics narrative from a source of widening economic disparity towards heightening workplace satisfaction and production efficiency. However, recent mass layoffs amidst installations of collaborative robots reveal the complexity of this balancing act (35).

The U.S. is in the midst of a major shift regarding how it shapes robotics advancement. Because a formal federal strategy does not yet exist, private and academic stakeholders are proactively drafting the blueprints themselves. Consolidating these ideas into a unified national policy will have far-reaching consequences for national security, geopolitical influence, labor markets, and manufacturing sovereignty. The world watches carefully as this shift unfolds.

Thanks for reading. For our next article we will be delving into China’s national robotic strategies – stay tuned.

Ellen H. Rumley – Policy Analyst

Allison Okamura – Vice President

IEEE RAS Science & Technology Watch Board

References

1. Boston Dynamics and Hyundai. School of Football | The Ghost Rabona | Boston Dynamics x Hyundai. YouTube, 2026.
2. University of Pennsylvania School of Engineering and Applied Science. “Penn and UMich Create World’s Smallest Programmable Autonomous Robots.” Penn Engineering Today, 2024.
3. “7 Cool NSF-Funded Robots That Are Advancing Science and Helping Society.” S. National Science Foundation, 7 Apr. 2021.
4. “AI 50 List: Top Artificial Intelligence Companies.” Forbes, 2026.
5. Atkinson, Robert D. A Time to Act: Policies to Strengthen the US Robotics Industry. Information Technology and Innovation Foundation (ITIF), 2025.
6. Association for Advancing Automation (A3). Policy Recommendations and Advocacy Principles for the U.S. Automate.org, 2025.
7. Consumer Technology Association. “Tech Sector Supports 18 Million US Jobs, Represents 12% of GDP, Says CTA.” CTA Press Releases, 2026.
8. Brookings Institution. Hardware and Software: A New Perspective on the Past and Future of Economic Growth. Brookings, 2024.
9. United States Department of Defense. Comptroller Budget Materials: Fiscal Year 2026. 2025.
10. Marrow, Michael. “DoD Plans Largest-Ever Investment in Drones, Anti-Drone Weapons.” DefenseScoop, 21 Apr. 2026.
11. Smith, Matt. “Humanoid Robots and Military AI Take the Field in Ukraine War.” CNBC, 30 May 2026.
12. Reuters Defense Bureau. “Iran Could Disrupt Strait of Hormuz with Drones Within Months.” Reuters, 4 Mar. 2026.
13. International Federation of Robotics. “US Robot Industry Returns to Double-Digit Growth.” IFR Press Releases, 2026.
14. McKinsey Global Institute. Ramping Up Manufacturing in America? McKinsey & Company, 2026.
15. International Federation of Robotics. “Japan Is World’s Number One Robot Maker.” IFR Press Releases, 2022.
16. Stanford University. Stanford Emerging Technologies Review 2026. 19 Feb. 2026.
17. National Science and Technology Council. Guidance for Implementation of National Security Presidential Memorandum 33 (NSPM-33). Jan. 2022.
18. White House Office of Science and Technology Policy. Guidelines for Foreign Talent Recruitment Programs. Feb. 2024.
19. United States, Congress, House. American Security Robotics Act. 119th Congress, H.R. 8189. Government Publishing Office, 2026.
20. Center for Strategic and International Studies. Innovation Lightbulb: Tracking CHIPS Act Incentives. CSIS, 2025.
21. Center for Strategic and International Studies. “Why the United States Needs Robots to Rebuild.” CSIS Strategic Technologies Blog, 2025.
22. Politico Pro Staff. “White House and Congressional Frameworks Shift Toward AI and Robotics.” Politico, 3 Dec. 2025.
23. United States, Congress, House of Representatives. “McGovern, Latta, Stevens, Obernolte Announce Re-Launch of Congressional Robotics Caucus.” Office of Congressman Jim McGovern, May 2025.
24. United States, Congress, House. National Commission on Robotics Act. 119th Congress, June 2026, H.R. 7334.
25. S. Army News Service. “Drone Dominance Program Receives First Order; Gauntlet II Gets Underway.” Defense Department News, 2026.
26. Congressional Budget Office. Cost Estimate and Analysis of the Golden Dome Air Defense Framework. May 2026.
27. S. Department of Energy. “Energy Department Advances Investments in AI and Robotics for Scientific Discovery.” DOE News, 2025.
28. Association for Advancing Automation. “A3 Policy Recommendations and Advocacy Principles for the U.S.” April 2024.
29. Special Competitive Studies Project. Memos to the President: National Robot Strategy. SCSP, 2025.
30. “Boston Dynamics Joins U.S. Robot Strategy Think Tank Led by Ex-Google CEO.” Seoul Economic Daily, March 2026.
31. Christensen, Henrik, et al. Robotics for a Better Tomorrow: 2024 US National Robotics Roadmap. UC San Diego / Academic Coalition, 2024.
32. Christensen, Henrik. Global Robotics Technology Roadmap 2025–2035. Independent Position Paper, 2025.
33. Appelbaum, Binyamin. “What Replaces Deported Immigrant Workers? Not Americans.” The New York Times, Feb. 2026.
34. Pew Research Center. “Key Findings About How Americans View Artificial Intelligence and Automated Systems.” Pew Short Reads, 12 Mar. 2026.
35. “Unions Furious as GM Replaces 1,000 Factory Zero Workers with 50 Robots.” Yahoo News, June 2026.

By Wanjiku Chebet Kanjumba, University of Florida

NASA is planning to send three small rovers to the Moon with a single instruction: Work out among yourselves how to explore a patch of ground.

The Cooperative Autonomous Distributed Robotic Exploration mission, or CADRE, will land on the side of the Moon facing Earth as part of NASA’s IM-3 launch, planned for late 2026. These rovers will spend roughly two weeks mapping the terrain as a self-guided team. No joystick will control them, and no human will approve each turn.

Engineers test whether the CADRE rovers can drive and coordinate on their own at NASA’s Jet Propulsion Laboratory in Pasadena, Calif. NASA/JPL-Caltech.

The rovers will elect a leader among themselves, assign their own tasks and redraw their plans as a group when one of them runs low on charge. If it succeeds, CADRE will be the first time NASA has operated multiple rovers beyond Earth as a single autonomous system.

That achievement will matter well beyond this mission, because NASA is scoping out future missions to the lunar South Pole, where water in the form of ice sits locked in craters that haven’t seen sunlight in billions of years. If researchers can chemically split that ice apart and turn it into propellant and breathable air, it could become the feedstock for a lunar economy built around fuel depots and life support made on the Moon.

Right now, the only demand for that ice comes from government contracts. And an operation that must be babysat from 239,000 miles (384,000 kilometers) away will not easily scale into a market. So, while a full-blown lunar economy is still far off, CADRE is testing whether robots can work unsupervised long enough, and in enough numbers, to keep a lunar operation running month after month.

I’m an aerospace engineering Ph.D. candidate researching guidance, navigation and control for spacecraft in-orbit servicing and active debris removal. I work on the same problem these rovers face: how a machine decides what to do next when it cannot call home for instructions.

Why Earth cannot drive

At the Moon’s South Pole, the terrain itself can cause communication disruptions. Commands reach a rover by way of relay satellites, and crater rims can block the line of sight to those relays. A rover that ventures down into a shadowed crater may lose contact for its entire trip.

Spotty communication doesn’t just make trying to drive annoying. It can cost the rover power it cannot recover.

A polar rover runs on a finite illumination budget. Solar panels charge the battery only while the Sun is up, and on the Moon sunrise is not a daily event. Night lasts about two Earth weeks, and at the poles, only a few ridges stay lit for long stretches. So for every minute a rover spends idle, awaiting new instructions, it is spending stored energy it cannot replace until the Sun comes back.

In the top image, a rover’s stored energy climbs while it charges in sunlight, then falls at the same rate whether it is working or waiting idly. In the bottom image, the gray bands are communication blackouts. A team waiting on commands from Earth stops until the link returns, while an autonomous team keeps assigning itself work, and the gap that opens between the two lines is work recovered from what would otherwise be dead time. Credit: Wanjiku Chebet Kanjumba.

Why the pole is the hard case

The lunar poles also come with unique challenges. Sunlight can swing from direct glare to absolute shadow as a rover drives down from a sunlit crater rim to the shadowed floor below, so a camera that worked at the top could go blind at the bottom.

The Moon also doesn’t have GPS satellites like Earth does, so rovers can’t know exactly where they are. They need to build their own maps from what their cameras and sensors see.

Temperatures inside permanently shadowed regions drop below minus 274 degrees Fahrenheit (minus 170 degrees Celsius). And the Moon’s dust is more dangerous than it sounds. An electric charge lifts it off the ground, and billions of years of tiny meteorite strikes have left every grain sharp and jagged.

That dust grinds at the wheel bearings and works past the seals. Keeping it out of the rover’s moving parts is still an unsolved problem. The dust can also film over the camera lenses that navigation depends on, leaving a rover unable to see where it is going or move safely.

Why one rover is not enough

Sending a single rover to check out a crater on the Moon is a risky mission. If it gets stuck, the campaign ends. If its instruments fail, no second machine can take the measurements. Multirobot teams distribute that risk and split up the jobs. One rover might carry the sensors and another the drill, while a lander positioned on a sunlit ridge acts as the power and communications hub.

Under a communications blackout, a rover team that waits for its commands from Earth has to stop. An autonomous team reassigns tasks among itself, selecting the next objective that it can reach and complete with the amount of power it has left. Dead time becomes work time.

In ground testing at NASA’s Jet Propulsion Laboratory, the CADRE rovers achieved this coordination. Faced with unexpected obstacles, they replanned paths as a group, and when one rover’s battery ran low, the whole team paused so they could continue together.

A prospecting team splits the work. A lander on the sunlit rim supplies power and relays communications, while a rover and a drill work the permanently shadowed crater floor below, where water ice may be trapped and no sunlight reaches. Once they drop past the rim’s radio horizon, they are out of contact and have to divide the tasks and manage their own power. Credit: Wanjiku Chebet Kanjumba.

What remains uncertain

A campaign at the lunar South Pole would need to run for months, and no robot team has yet worked that hard, for that long, that far from help. Engineers are developing ways for rovers to navigate without satellite positioning and to make decisions onboard, but that software still has to be tested on the surface.

Meanwhile, as of mid-August 2026, China’s Chang’e-7 mission is waiting at Wenchang, the launch site on Hainan Island in China, with liftoff expected by the end of the year. Its lander, rover and hopping probe are meant to work at the pole as a team, with the probe built to leap into permanently shadowed craters. If the U.S. wants to keep pace, it will need its own robots that coordinate without supervision.

Finally, there is no agreed way for a rover built by one company to hand a task to a tool built by another, or to decide which one works the crater floor first. As autonomous rover teams improve, the groups building them will have to work out those rules.

CADRE will tell scientists whether three small rovers can reason together on the Moon. The harder question is whether a dozen different machines from different builders can do the same thing, reliably, for years.

Wanjiku Chebet Kanjumba, Ph.D. Candidate in Aerospace Engineering, University of Florida

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.

By Emmet Cole

Robots clean rivers and sort waste, monitor ecosystems, and inspect renewable-energy infrastructure. But even the greenest robot has an environmental footprint.

If robotics is going to help build a more sustainable world, the robotics community has to answer some potentially awkward questions, starting with this one: How sustainable are robots themselves?

Across a full lifecycle, from rare earth mineral extraction and manufacturing to operation and end-of-life, robots have an environmental impact. But the robotics community has, until now, lacked dedicated tools for calculating it.

The Robotics Eco-Label project, led by Bram Vanderborght at Belgium’s Vrije Universiteit Brussel, is an attempt to address that gap with a lightweight, web-based Toolkit that provides roboticists with a way to quantify robot sustainability.

Separating a robot’s core technologies into materials, energy sources, sensors, processors, actuators, design, and recyclability, the Robotics Eco-Label Toolkit then evaluates each based on five metrics: resource conservation, lifecycle extension, carbon footprint, energy efficiency, and circularity. The numbers are combined in a weighted matrix to yield a 0–100 Eco-Score for Robots. The weights used are not currently fixed; that will be one of the targets of further research and collaboration. (For an indicative score on your robot, try the project’s interactive self-assessment tool here.)

A low score in one area might point to energy-hungry actuators, limited repairability, hard-to-recycle materials, or a lack of end-of-life planning. A higher score, by contrast, suggests that sustainability has been considered across the system. One of the project’s stated goals is to make environmental trade-offs visible early enough in the development process to shape sustainable robot design choices.

The project also includes educational content and community features so researchers and developers can compare approaches, share case studies, and turn broad sustainability goals into improved design decisions. For companies, Eco-Label could well turn out to be a way to achieve competitive advantage, while it could also help buyers make more informed decisions.

Robotics Eco-Label is just one of the IEEE RAS Sustainability Grant-funded projects showcased at ICRA in June. The grants are a key component of IEEE RAS’ broader effort to make sustainability a more visible part of robotics research, education, design, and deployment.

Are you buying more robot than you need?

Sometimes, sustainable robotics starts with better purchasing decisions. This includes avoiding overspecification; that is, buying robots that are larger or more capable than necessary.

Matching robots to their intended workload can reduce unused capacity, avoid unnecessary material use, and cut wasted energy over a robot’s life cycle.

That principle underpins the work of an IEEE RAS-funded team, led by Antun Skuric, that has developed an open-source platform for assessing the sustainability of collaborative robots.

To use it, you define a required workspace, payload, and trajectory, and the platform identifies the minimum-mass robot that satisfies your application requirements.

The application features interactive tools that enable users to visualize and jog the robot, inspect task-related variables and requirements, view the reachable space, and observe important robot configurations. This enables robot performance to be calculated based on specific task conditions and requirements.

Can robots really help communities overcome energy poverty?

Energy poverty and inefficient solar energy utilization are major challenges in Nigeria, where more than 90 million people lack reliable electricity access.

The SolarPeer 360 project, led by Umar Adetola Abdulganiyy, a student at the Federal University of Technology, Minna, Nigeria, takes on this challenge through a combination of robotic solar tracking, AI-assisted optimization, and peer-to-peer energy distribution.

A timely reminder that robots can help support sustainability goals directly, SolarPeer addresses two connected problems, inefficient small solar installations and the lack of transparent, affordable mechanisms for sharing surplus renewable energy among households and small businesses.

According to the team, SolarPeer 360 is built on a sustainability logic in which “energy captured more efficiently can be shared more fairly, and energy shared more transparently can create local economic value while reducing waste and fossil-fuel dependence.”

SolarPeer links smart solar capture through tracking, controlled and metered distribution through embedded electronics, and behavior optimization through data, interfaces, and AI guidance.

Early results indicate meaningful gains: The team reported a 60.3 percent gain in average power in one tracked-versus-fixed solar comparison, with measured average power rising from 3.83 W to 6.14 W. Meanwhile, AI-guided energy advisory and optimization contributed to a ~25 percent reduction in energy wastage in the testing environment.

Beyond the lab, the team deployed five community mini-systems and ran a solar training and empowerment workshop that reached more than 500 students.

Can Caretta work faster?

Sustainability in robotics is not just a technical challenge. It’s also a cultural and educational challenge for the next generation of engineers, researchers, teachers, and users.

That’s part of the reasoning behind the ‘Caretta’ project, led by Mustafa Kemal Ambar, which brought robotics and sustainability education to students aged 8 to 16 on the island of Cyprus.

Inspired by the Caretta sea turtle, the project produced a functional robot prototype designed to reduce coastal pollution. Two successful coastal clean-up events were held and more than 30 students were engaged in the project through seminars and hands-on learning.

Treating the beach as both a test site and a classroom, students saw how engineering connects to local environmental problems and community needs.

During a field exercise, one student pointed to plastic debris near the water and observed: “Maybe turtles won’t eat this anymore if Caretta works faster.”

Can the robotics community work faster to build sustainability into its foundations and practices? Early results from IEEE RAS Sustainability Grant projects suggest that work is already underway.

Want to learn more?

RAS University now has a free new class on Sustainable Robotics, you can learn more here.

You can also follow activities from the Sustainability and Climate Change Committee, including upcoming grant calls here.

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.

Robots have been a prolific theme in Japanese pop culture and media since the 1950s, which includes global icons like the Transformers, Astro Boy, and Doraemon (1). Perhaps not coincidentally, Japanese citizens have a positive outlook on robotic technologies and their use in the labor sector compared to many other nations (2); much to their advantage, as robotics continues to be a vital pillar for ensuring Japan’s economic resilience and future growth. Here, we discuss priorities and long-term agendas as presented in Japan’s national robotics strategies.

Natural disasters, aging, and economic competition

The combination of its geographic location, geologically active landmass, and topography renders Japan prone to frequent natural disasters. Although weather control is among Japan’s lofty research goals for the next 25 years (3), geological disturbances remain inevitable. As such, disaster mitigation and response are high priorities that serve as motivation for robotics development in Japan (4). Intelligent machines are expected to expedite disaster response and search-and-rescue, while minimizing exposure of workers and volunteers to unnecessary hazards – a notable reminder being the 2011 Fukushima Disaster (5).

Aging workforce is a second pressing issue. Japan is currently the oldest demographic nation, defined as of 2007 as a “super-aged” society with nearly a third of citizens aged 65 or older (6). According to government roadmaps (4), Japan views robotic development as a crucial tool for preventing a GDP collapse due to the rapid labor shortage, specifically in services like healthcare, elderly assistance, and manual labor. Unlike countries which opt for immigration reform for addressing fluctuation in labor markets, Japan maintains historically tight immigration regulations. A strong preference towards preserving traditional values may influence their massive investments towards shifting to a robotic workforce (7).

Japan has a mature robotics industry, claiming the title of the largest manufacturer and exporter of industrial robots for the last half century (8). 45% of the global supply originated from Japan in 2022, with nearly an 80% export rate to countries such as EU members, the USA, and China (9). However, the combination of economic decline and competition from oversea manufacturing present the challenge of maintaining leadership in the global value chain. Leveraging their specialty in robotics hardware design is crucial to maintain a competitive edge in the global economy.

Society 5.0

Japan is betting on bold technological solutions for ensuring the prosperity of their future society. To do so, the government proposes a streamlined effort towards technocentric “super-smart cities” (coined Society 5.0) (10). The directive is set by the Science and Technology Basic Plan, a large-scale roadmap updated every five years (with 2026 marking the transition into the 7th Plan). An additional Integrated Innovation Strategy is used as an annually-updated agile document for integrating urgent priorities into the larger Basic Plan.

Super cities

Japan is investing heavily into research on human-robot interactions, so that users are embedded early on into technological design considerations of Society 5.0. Accordingly, in 2020 the Japanese Parliament passed the Super City Amendment, which allows regulatory sandboxes for specific “super cities” (currently Osaka and Tsukuba) to bypass certain laws that otherwise impede the integration of technologies like AI-driven robots into the public space (11, 12). These large-scale tests are carried out preceding what they hope could become widespread implementation of robotics to mitigate issues such as labor shortage. Following suit, as of late September 2025, the automobile industry Toyota has launched their own version of a Super City, Woven City, which provides a campus for employees to work, live, and even raise a family (13).

Moonshot programs

Japan also utilizes high-risk, high-impact research roadmaps with 30-year timelines, called Moonshot Programs, to encourage bold technological solutions for the future (14). The authors of these programs are a blend of members from government, industry, and academia. Two of the current ten Moonshot Goals are directly related to robotics.

Goal 1 of the Moonshot Program is to Overcome Limitations of Body, Brain, Space, and Time, a goal largely focused on the creation of Cybernetic Avatars – the fusing of sensations and motions between human users and tele-operated robots. By 2050, they hope to create the technology and infrastructure to implement Cybernetic Avatars into everyday society. Goal 3 of the Moonshot Program is the Co-Evolution of Robots and AI. Sub-goals within Goal 3 include the creation of robots that 1) 90% of citizens will feel comfortable engaging with by 2030, 2) work in remote and dangerous locations by 2050, and 3) autonomously innovate by 2050.

Humanoids and physical AI

In December 2025, a strategic update has been made to Goal 3 of the Moonshot Program, which prioritizes the development of general-purpose autonomous humanoid robots which can adapt to human-centric environments (15). Japan had in fact led humanoid research decades back (16), already developing Honda’s famous Asimo in 2000 (17) – but despite these early successes, humanoids remained astronomically expensive to manufacture, impractical for real-world applications, and remained publicity tools for promoting other industrial products. These commercialization challenges have largely been overcome with advancements in low-cost and back-drivable actuation components, additive manufacturing, and generalizable AI. As countries like China and the USA have sparked a humanoids build-up race, Japan is playing catch-up to re-assert themselves in this arena, which is rapidly taking center stage in the tech world (18).

To keep up in this humanoids race, it is crucial for Japan to invest heavily in domestic AI. In line with this need, in the same month as the Moonshot Program update, the Japanese Ministry of Economy, Trade and Industry (METI) decided to quadruple its previous investments in AI to $8 billion USD (now the third highest expenditure for AI just after China and the USA), with approximately $2.5 billion USD of this fund allocated specifically for physical AI (robots) (19).

In March 2026, Japan has also included updates to their Integrated Innovation Strategy (20), with plans to install R&D hubs across its 16 sectors for training domestic AI-robots and recruit both local and international talent. Overall, it would seem that Japan is leveraging its expertise in precision hardware to perform large-scale physical training for high-precision AI-driven robots.

Strategic relations with the USA

For the better part of the last century, Japan and the USA have been allies and trading partners. Japan’s recent Moonshot Program, along with their National Defense Strategy from 2022, both state the importance of their continued partnership (21). Notably, the USA is a global leader in AI but lacks a government-backed robotic hardware strategy, which could ensure a symbiotic relationship with Japan’s mature hardware ecosystem (22).

Supply chain pressures have also incentivized Japan and the USA’s joint collaboration to conduct deep-sea rare-earth mining using advanced robotic systems off the coast of Japan, in line with Goal 3 of the Moonshot Program. The first successful mining test, which involved underwater autonomous vehicles, was conducted in February 2026 (23). The use of robotics for underwater monitoring will also surely become critical for monitoring the impact of deep-sea mining to the marine environment, a highly contested topic (24).

In a historical move amidst mounting geopolitical tensions, Japan has also recently introduced dual-use technologies as a priority for their revised 7th Basic Plan (25), which demands the ramp-up of domestic R&D and research on unmanned vehicles and drone technologies (a move strongly endorsed by the USA). This decision comes with the high expenditure commitment of 2% of the national GDP.

Japan has ambitious societal plans for the coming decades, and robotics will play an increasing vital role – from sustaining domestic services, to enabling the physical autonomy of citizens, and revitalizing the economy. Ultimately, Japan’s roadmaps reveal a future where robots are partners for shaping a resilient and human-centric society.

On that note – our next month’s article will feature robotics strategies from the USA. Thanks for reading, and stay tuned.

This is a continuation of a monthly series on international robotics roadmaps. We welcome questions, comments, and feedback for future editions at ras@ieee.org.

Ellen H. Rumley – Policy Analyst

Allison Okamura – Vice President

IEEE RAS Science & Technology Watch Board

References

[1] Hornyak, T. N. (2006). Loving the Machine: The Art and Science of Japanese Robots. Tokyo: Kodansha International.

[2] Mitsubishi Research Institute. (2022). How different countries perceive robots: An original survey.

[3] Cabinet Office, Government of Japan. (2021). Moonshot Goal 8: Realization of a society safe from the threat of extreme winds and rains by controlling and modifying the weather by 2050. Bureau of Science, Technology and Innovation.

[4] Cabinet Office, Government of Japan. (2025). Integrated innovation strategy 2025. [5] Encyclopaedia Britannica. (2026, March 9). Fukushima accident.

[6] Cabinet Office, Government of Japan. (2024). Annual report on the state of the formation of a resilient society for an aging population: White paper on the aging society 2024.

[7] Speed, J. Japan aims to take in 1.23M foreign workers under labor migration programs. The Japan Times.

[8] Nasdaq. (2021, October 25). Japan’s robot dominance.

[9] International Federation of Robotics. (2022, March 10). Japan is world’s number one robot maker.

[10] Cabinet Office, Government of Japan. (2016). The 5th Science and Technology Basic Plan (2016–2020).

[11] Cabinet Office, Government of Japan. (2020). Act partially amending the Act on National Strategic Special Zones and the Act on Special Districts for Structural Reform (Act No. 34 of 2020).

[12] Cabinet Office, Government of Japan. (2022). Designation of the Super City type National Strategic Special Zones: Osaka Prefecture/Osaka City and Tsukuba City. Bureau of Regional Revitalization.

[13] Toyota Motor Corporation. (2025, September 25). Toyota Woven City officially launches as a test course for the future of mobility [Press release].

[14] Japan Science and Technology Agency. (n.d.). Moonshot R&D | TOP.

[15] Kuniyoshi, Y. (2025, October 28). Appendix 3: PD’s policy (Moonshot Goal 3). Japan Science and Technology Agency (JST).

[16] Waseda University. (2026, May 8). The robots of Waseda: A 50-year journey in humanoid innovation. Waseda University News.

[17] Honda Motor Co., Ltd. (n.d.). History of robotics development. Honda Global Corporate Website.

[18] Morgan Stanley Research. (2025, May). Humanoids: A $5 trillion market. Morgan Stanley Investment Management.

[19] Nohara, Yoshiaki, and Komaki Ito. “Japan to quadruple spending support for chips and AI in budget.” The Japan Times, 26 Dec. 2025.

[20] The Japan News. (2024, May 20). Japanese govt to set up global hubs for AI robotics research.

[21] Cabinet Secretariat. (2022, December 16). National security strategy of Japan [Provisional translation].

[22] Fujitsu. (2025, October 3). Fujitsu expands strategic collaboration with NVIDIA to deliver full-stack AI infrastructure [Press release].

[23] The Japan News. (2024, August 24). Japan deep-sea drilling ship Chikyu extracts rare earth mud; represents major breakthrough for domestic supply. The Japan News. [24] Lelyveld, M. (2024, September 10). Japan’s deep-sea gamble: A new rare earth frontier in the Pacific. The Diplomat.

[25] Cabinet Office, Government of Japan. (2026, March 27). The 7th science, technology, and innovation basic plan (FY2026–FY2030) [Cabinet Decision].

Scientists from Nanyang Technological University, Singapore (NTU Singapore) have developed a tiny seed-sized robot that can navigate across soft and uneven surfaces to perform five surgical functions wirelessly, paving the way for developing robots to make surgeries and medical treatments more precise.

The miniature robot, measuring just 4.4 mm in length and controlled by weak magnetic fields, can move, cut biological tissues, release drugs, grip and store tissue samples, or generate heat remotely at any one time. It takes under a second to switch between these functions.

Led by Associate Professor Lum Guo Zhan from NTU’s School of Mechanical and Aerospace Engineering (MAE), the work was recently published in the academic journal Advanced Materials.

Using magnetic coils in the laboratory to remotely control the robot, the team was able to make the robot deploy different tools and perform different functions, such as activating a tiny blade to cut through tissue, or emitting heat to a targeted area, which could be relevant for approaches being studied that use heat for cancer treatment.

“Most magnetic robots like this can perform only one or two functions. Our latest invention can now do five, and our long-term goal is for doctors to use these mini robots in the body, navigate them to a targeted location, and use them to perform treatments,” said Assoc Prof Lum, who is a pioneer in miniature robots made from soft, flexible materials.

Mini robots are being studied worldwide as a possible way to make minimally invasive surgical and medical procedures safer, less painful and more precise.

Such devices could one day allow doctors to carry out targeted surgeries deep inside the body without large cuts or bulky surgical instruments.

Overcoming a key robotics challenge

To fit multiple functions into a robot only a few millimetres long, the NTU team developed a device for controlling movements that is activated by magnetic fields and which can be reprogrammed in under a second.

The robot is made from soft magnetic materials, including PDMS and Ecoflex, which are silicone-based materials commonly used in soft robotics as they are flexible and can be shaped into small structures.

These materials are embedded with magnetic microparticles measuring 5 micrometres each, allowing different parts of the robot to respond to magnetic fields.

At the centre of the device is a magnetic module that can be magnetised, demagnetised and remagnetised in different directions.

Each magnetic orientation activates a different function of the robot, allowing the same mobile robot to perform five different functions, including cutting and grasping tissues.

The researchers also engineered different regions of the robot to ensure that only one part, but not the rest, responds to the same magnetic field.

This means that only one part of the robot reacts to a magnetic field to change its shape to activate a tool or function, while other parts remain still and unchanged in their current forms, addressing a major limitation in miniature magnetic robots.

At small scales, magnetic fields often affect the entire device at once, causing it to behave like a single magnet, with all parts reacting to a magnetic field, thus limiting how precisely it can move or activate different tools.

Most miniature magnetic robots are also limited to five degrees of freedom. They can only move along three axes and rotate in two directions.

The NTU robot adds a sixth movement, rolling, which allows it to spin around its own long axis. This gives the robot finer control over how it can be positioned, which is important for navigating narrow, soft and irregular spaces, such as those inside the body.

Unlike slime-like mini robots, the NTU robot has a solid but flexible body, making it sturdier and easier to retrieve after use.

Tested on biological tissues

The NTU team tested the robot’s surgical functions using biological tissue models, including chicken liver, as well as gelatin-based materials that simulate soft tissue.

In laboratory tests, the robot cut through biological tissues, dispensing particles simulating drug particles, gripped and stored tissue samples, and generated localised heat after being induced by magnetic fields.

To produce heat, the researchers exposed the robot to a high-frequency alternating magnetic field. This caused magnetic materials inside the device to generate heat remotely, in an approach relevant to magnetic hyperthermia methods being explored in cancer treatment.

The team also evaluated the biocompatibility of the robot’s materials by exposing them to human skin cells under laboratory conditions.

More than 99 per cent of the cells remained viable after exposure to the robot’s materials, similar to the control group, suggesting that the materials were largely non-toxic under the experimental conditions.

The team – including NTU’s MAE alumnae Dr Chelsea Shan Xian Ng and Yu Xuan Yeoh, and current PhD student Nicholas Yong Wei Foo, who are co-authors of the research – is now exploring how future versions could be combined with imaging technologies, sensing systems and clinically realistic artificial organ models that better mimic the physical behaviour of human tissues.

Assoc Prof Lum is also working with surgeons to understand how mini robotic systems could eventually fit into real clinical workflows.

“For these robots to move closer to practical use, we need to understand not just how they work in the lab, but how they could be guided, monitored and controlled in realistic medical settings,” he added.

Giving independent comments, Dr Yeo Leong Litt, Leonard, Senior Consultant from the Division of Neurology at the National University Hospital, said: “These millimetre-scale magnetically guided robots are truly remarkable in their ability to traverse complicated environments and then perform a variety of tasks such as deliver medication to a location, perform biopsies and administer therapeutic heat remotely. I can envision that they have the potential to replace many aspects of interventional radiological surgery and become a new mode of therapy in medicine.”

Read the work in full

Miniature Soft Robot With Magnetically Reprogrammable Surgical Functions, Chelsea Shan Xian Ng, Yu Xuan Yeoh, Nicholas Yong Wei Foo, Keerthana Radhakrishnan, Guo Zhan Lum.

By Alex Shipps

Robots walking down the street, surrounded by astounded onlookers, is an increasingly common sight. But these machines aren’t yet the do-it-all assistants you’d want working in a kitchen or factory, and a major bottleneck is data. Much like humans, robots learn best by experience. The challenge is that it’s labor-intensive and time-consuming to physically teach these machines so many actions across different settings.

“One natural idea is to use simulation as a training ground. While there has been significant progress over the last few years in the physics engines that power robotics simulators, one of the remaining challenges has been creating sufficiently rich and diverse simulation content to capture the complexity of the real world,” says Russ Tedrake, the Toyota Professor of Electrical Engineering and Computer Science (EECS), Aeronautics and Astronautics, and Mechanical Engineering at MIT, and a principal investigator at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).

It turns out that AI agents, or semi-autonomous programs that “think” and complete well-defined tasks, could help produce the lifelike virtual settings that robots need. The new “SceneSmith” system developed by researchers at MIT CSAIL and Toyota Research Institute uses three agents to piece together the objects, walls, and overall look of a 3D scene. Its recreations of indoor spaces such as restaurants, bedrooms, and hotels are more realistic and detailed than prior systems, helping robots practice skills and try out different ways of doing tasks before they’re powered on. In turn, engineers save time on real-world testing.

The agents have a sense of how everyday places are supposed to look because they each call on a multi-modal system called a vision-language model (VLM), specifically the state-of-the-art VLM GPT-5.2. It’s trained on lots of text and images from the internet to handle more visual prompts. This advanced model gives each agent a sort of spatial knowledge: First, a “designer” agent generates the elements of a scene, then a “critic” advises whether it looks realistic, and finally, an “orchestrator” manages their back-and-forth, deciding when the design is done. Once the three VLMs wrap up their creative collaboration, the scene is ready to load directly into physics simulation software.

“We’ve found that the system can construct 3D scenes the way a human designer would,” says MIT EECS PhD student Nicholas Pfaff, a CSAIL researcher and a lead author on a paper with Tedrake presenting the work. “We made over 1,300 scenes using a leading VLM that has internet-scale priors, and it made insanely creative and diverse arrangements. I hadn’t taught the system to do that in the prompts; it just improvised.”

Talk to my agent

Thanks to VLM agents, you can ask SceneSmith to do things like “generate a garage with a car, a workbench, tires stacked in the corner, and a ladder against the wall,” and get a virtual playground rich with objects a robot can tinker with. These rooms are decorated with up to six times more items per scene than prior methods, making them great for helping robots learn skills such as putting a cup in the sink, placing fruit on plates, and moving a soda can from a shelf to a table.

With so many rich virtual environments handy, you can evaluate whether your robot is ready for deployment without so much trial and error in the physical world. The researchers tested out different action plans (also called “policies”) in SceneSmith’s digital worlds, generating 100 unique spaces in the process. A VLM agent evaluated each attempt, and it found the robot’s plans were faulty, with the machine often failing at its chores. Humans agreed with the model’s verdicts over 99 percent of the time, which could help roboticists weed out flawed approaches in simulation before a robot moves in the real world.

But how realistic are these virtual worlds, really? It can be difficult to prove outright, so the researchers approached the question from several angles. The most telling test: they dropped a pretrained robot policy — an AI controller trained largely on real-world data, which had never seen a SceneSmith scene — into the generated environments. In one test, users told the system to “take the apple from the bowl and place it onto the cutting board,” and the simulated robot did exactly that. If the scenes didn’t closely resemble the real settings the policy had learned from, it simply wouldn’t have worked.

The team also teleoperated robots through the virtual spaces, guiding them to open cabinets, put away bottles, and navigate between rooms. Their experiments revealed that the environments hold up under sustained physical interaction, expanding beyond visual inspection.

Behind the scenes

The agents that SceneSmith uses each have a well-defined role in the generative process, fleshing out scenes in stages. They essentially create a floor plan and bring it to life.

Let’s say you wanted to create a scene similar to the first floor of a house. The “designer” VLM would start with a general layout, which the “critic” reviews, and then the “orchestrator” signs off. The agents repeat this approach for each step: adding furniture, placing objects on walls and then ceilings, and finally, dropping in objects that robots can manipulate. For example, the VLMs can add cabinets that the robots can open and close — an articulated item, which prior baselines didn’t often have.

At each stage, the second VLM ensures the scene is practical, advising that a bathtub is removed from a living room, for example. The third VLM ensures a high-quality scene is generated, even taking the design process a few turns back if the visuals aren’t up to par. Once the three VLMs wrap up their creative collaboration, the mechanics of the physical world are added via simulation software.

With a sound understanding of how rooms should look, where objects should be placed, and real-world physics, SceneSmith has a noticeable edge over prior methods. Compared to scene-generation baselines such as “HSM” and “Holodeck,” SceneSmith made environments with more objects, including a private office, a pottery store, and even a Minecraft-themed gaming room.

SceneSmith was also a favorite among over 200 users. They found the system’s visuals to be more realistic over 90 percent of the time. They also observed that, generally speaking, it followed prompts more closely than other approaches did. In other words, it was the best at generating the virtual playgrounds users actually wanted to see.

A system of many talents

Realism, diversity, and richness are all strong suits for SceneSmith, even when it comes to generating individual 3D objects. You can prompt it to create a rolling serving cart, and it’ll make a 2D image that it then turns into a detailed model with physical properties like mass, friction, and inertia.

Such a detailed process does come with a speed trade-off, though. It can take multiple hours to produce a single scene because the agents are creating and closely scrutinizing each object. With more computing power, the system could see dramatic increases in efficiency. CSAIL engineers are also hoping to expand to deformable objects (like sponges), should extensive 3D libraries become available.

“SceneSmith represents a significant advance in this regard by providing an agentic framework for generating simulation-ready indoor environments just from a simple text prompt,” says Jeremy Binagia, an applied scientist at Amazon Robotics who wasn’t involved in the research. “It advances the state of the art in several ways, including pushing the limits of the density of objects in the simulated environment, ensuring that all of the objects are physically accurate (versus just being visually realistic), and creating assets that are not constrained to a fixed library, since they can be generated via text-to-3D.”

Pfaff and Tedrake wrote the paper with Thomas Cohn SM ’24, an MIT PhD student and CSAIL researcher; and Toyota Research Institute roboticists Sergey Zakharov and Rick Cory SM ’08, PhD ’10. Their work was supported, in part, by Amazon, the U.S. Office of Naval Research, the Toyota Research Institute, and the U.S. National Science Foundation.

The team presented their findings as a spotlight at the 2026 International Conference on Machine Learning.

By Emmet Cole

Researchers at the University of Leeds and collaborators from the University of California San Diego won the Best Paper Award at RoboSoft, the leading international conference focused on soft robotics research. Soft robotics is gaining attention in medical applications because compliant machines can interact more safely with delicate objects and complex anatomy.

The award-winning paper describes a 1.8 mm soft growing robot that can be steered magnetically, sense its own shape in real time, and operate without internal pressure. These advances could help improve patient outcomes following minimally invasive procedures.

We spoke with lead author Benjamin Calmé about the team’s work.

Q: Congratulations to you and your team on winning the award. Before we get into the paper itself, could you tell readers a little about yourself and how you arrived in this field?

Benjamin Calmé: My path into robotics is a bit unusual. I’m from France and I actually started with a medical degree. Over time I realized I was more interested in research than in day‑to‑day clinical practice. In France, if you want a research and teaching career in medicine, you’re expected to get an equivalent engineering degree as well. That’s what pushed me toward engineering—and from there into robotics.

That led me into robotics labs in Paris and Strasbourg, where I worked on medical robotic platforms for applications such as needle insertion inside MRI scanners and technologies to help runners reduce injury risk. Through those projects I fell in love with robotics.

Staying in medical robotics was a natural extension of my previous work. It’s also why I was hired in Leeds. That is, you need someone who can translate between surgeons and engineers. Surgeons know what they want clinically, engineers know how they want to solve the problem, and those views don’t always match. My role is often to say, “This is what the surgeon actually wants, and this is how we can realistically help.”

I also help when systems move toward pre-clinical testing: designing study protocols, discussing workflows with clinicians, and helping them understand how a new platform behaves compared with conventional tools.

Q: For readers outside soft robotics, what is a “growing robot”?

Benjamin Calmé: It moves more like a plant than a traditional robot. Instead of pushing or dragging its whole body through the environment, it extends at the tip by bringing material from inside the robot outward. In effect, the robot grows into the space ahead of it.

We often call these ‘vine robots’ because the idea is inspired by climbing plants. The body is soft and compliant, so when it encounters obstacles, it can deform and follow paths of lower resistance.

That can be very useful inside the body, where space is constrained and tissues are delicate. Rather than forcing its way forward, the robot can adapt to the environment.

Q: What problems does this work address?

Benjamin Calmé: A central issue is friction. In many procedures, conventional flexible tools still rub against tissue as they are inserted and withdrawn. That can cause irritation or inflammation.

With a growing robot, the body sections already in place move far less because new material advances at the tip. That can significantly reduce friction along the path.

For patients, that can mean less discomfort and fewer side effects. It can also help clinicians attempt procedures in anatomical regions, such as the brain, where the margin for error is very small.

Q: Your paper highlights pressure-free growth. Why is this important?

Benjamin Calmé: Safety is one reason, and controllability is another.

Many earlier growing robots relied on internal air pressure. But if you are working in fluid-filled spaces near the spine, or inside blood vessels, introducing air because of a leak is unacceptable. By removing the need for internal pressure, that risk disappears.

There is also a practical control benefit. In some earlier prototypes, we used pressure to grow and magnetic fields to steer. That meant alternating between growth and steering steps rather than doing both together.

Now we can grow and steer simultaneously, which makes the system faster and more usable.

Q: What is novel about your approach compared with previous systems?

Benjamin Calmé: One key contribution is combining shape control and shape sensing in a structure that can still be miniaturized. Our current prototype has an outer diameter of just 1.8 mm. Many previous designs place separate actuators or sensors along the robot body. Those solutions can become slow, bulky, or difficult to shrink to catheter scale.

We instead embed magnetic functionality directly into the silicone body by mixing magnetic particles into the material, molding it, and then magnetizing regions in controlled directions.

You can think of that pattern as the robot’s magnetic DNA. When we apply an external magnetic field, the robot bends into predictable shapes. By tracking how those magnetized regions move, we can also estimate the robot’s shape in real time. So, with one integrated structure, we achieve both actuation and sensing.

Q: How does the real-time shape control system operate?

Benjamin Calmé: We first manufacture the internal tail section of the robot from silicone containing magnetic particles. That tail later everts and becomes the outer body as the robot grows.

Before that happens, we place the material in a coil and apply a strong magnetic field. By orienting the material carefully during magnetization, we assign different magnetic directions along its length and cross-section.

Once deployed, moving an external magnet around the robot creates specific deformations. Because the magnetic pattern is known, we can also infer shape as the robot moves, with sensing updates up to 500 Hz.

A subtle challenge was preventing unwanted attraction between inner and outer layers, which would increase friction. Designing patterns that gave useful control without causing sticking required significant optimization.

Q: The paper also demonstrates retroflexion and biome sampling in an ex vivo stomach model. Why are those meaningful milestones?

Benjamin Calmé: Retroflexion and biopsy in an ex vivo stomach are important because they show the robot can perform endoscopic maneuvers in a realistic anatomy, not just in bench-top tests. Retroflexion proves it can safely reach difficult angles without the high friction and tissue stress of conventional scopes, and successful sampling shows it can precisely position tools and carry out a core clinical task. Together, they’re a first demonstration that this pressure-free growing robot can do meaningful work in settings that resemble actual medical procedures.

Q: What were the biggest engineering hurdles?

Benjamin Calmé: Manufacturing at this scale was a major challenge. Our catheter has an outer diameter of about 1.8 mm, with wall thickness near 100 microns.

That required careful control of injection molding, vacuum processes, particle distribution, and defect prevention. Tiny bubbles or inconsistencies can affect both mechanical performance and magnetic behavior.

Q: What comes next, and how close is clinical use?

Benjamin Calmé: We are especially interested in neural and spinal applications, where precise placement of electrodes could help restore function after injury.

We are still early in development. A realistic near-term target is robust pre-clinical performance: safety, biocompatibility, and successful in-vivo demonstrations.

Clinical adoption takes much longer because regulation must be rigorous. That is appropriate when patient safety is involved.

Q: How would you explain this advance to a non-technical person?

Benjamin Calmé: I sometimes use the image of brain surgery done with chopsticks. Right now, your surgeon is often working around the most sensitive parts of your body with rigid tools that must be manipulated with extreme care. We’re trying to replace those chopsticks with a softer, more precise, less dangerous tool.

This soft growing robot can be very small and dexterous. It can reduce some types of human error, like tremor, and it doesn’t take much space, so there’s more room for other instruments and better imaging. Surgeons can also see and reach regions that used to be in blind spots.

In the simplest terms, we’re developing a new class of soft, growing robotic tools that can sneak into delicate spaces in the body, minimize damage along the way, and give surgeons more control and information than they have with today’s rigid instruments.

The paper, “Pressure-free Magnetic Soft Growing Robot with Real-Time Shape Control and Sensing for Biome Sampling,” appears in the proceedings of the 2026 IEEE 9th International Conference on Soft Robotics (RoboSoft).

This article originally appeared on IEEE RAS.

A natural question is whether we can replace some of this real-robot work with a simulator. Classical physics-based simulators are powerful, but building one for a new task means manually modeling geometries, contacts, friction, and deformation, and the resulting simulator often still does not match reality closely enough for policies trained inside it to transfer.

In our work, we take a different route. We build an Interactive World Simulator: a learned, action-conditioned video prediction model that, given the current image and a sequence of robot actions, predicts the next frames purely in pixel space, with no physics engine inside. You can plug in a teleoperation device and control the robot through this learned world model for more than 10 minutes at 15 FPS on a single RTX 4090, and the predicted video stays stable and physically plausible.

The key idea is that, if the simulator is faithful enough, we could unlock two long-standing bottlenecks in robot learning:

– Data generation for training becomes cheap, because we can collect demonstrations inside the simulator.
– Policy evaluation becomes scalable and reproducible, because we can roll many policies through the simulator under identical conditions.

What the world simulator can do
We trained our world simulator on four manipulation tasks that span very different physical regimes: T pushing (rigid-body contact), rope routing (deformable–rigid interaction with a clip), mug grasping (fine-grained gripper dynamics), and pile sweeping (manipulating piles of objects). All four behaviors are learned from interaction data alone, with no physics priors hard-coded.

A few examples of what the model captures:

Rope routing: it correctly distinguishes between the rope actually being inserted into the clip and the rope swinging past it without making contact. Crucially, it does not bias toward either outcome — it follows what the actions imply.

Mug grasping: it captures fine-grained effects such as the mug slipping out of the gripper, or the handle being nudged and rotated.

Pile sweeping: it can generate video for multiple viewpoints consistently

You can try this directly on our project page: just open a browser and play with it using your keyboard!

How we built the interactive world simulator

At a high level, Interactive World Simulator is trained in two stages. First, we trained an autoencoder that compresses RGB images into compact 2D latent representations and reconstructs them back into images. This lets the model reason in a lower-dimensional space while still producing high-fidelity pixel-level outputs.

Second, we froze this autoencoder and trained an action-conditioned dynamics model in the latent space. Given past visual latents and robot actions, the model predicts the next latent state, which is then decoded back into an image. At inference time, this process is repeated autoregressively: predicted frames become part of the context for predicting future frames. Because prediction happens in latent space and uses consistency models, the simulator can run interactively while remaining stable over long horizons.

How is it different from prior works?

This differs from several existing approaches to robot simulation and world modeling. Classical physics simulators can be powerful, but they often require manually specifying object geometry, contact dynamics, friction, and deformation, and the resulting simulation may still have a large sim-to-real gap. Recent video-generation and world-model methods offer a more data-driven alternative, but many are either not explicitly conditioned on robot actions, too slow for real-time interaction, closed-source, or unstable under long-horizon rollouts.

In contrast, our Interactive World Simulator is action-conditioned, produces physically accurate pixel-level predictions, and supports stable interactions for more than 10 minutes at 15 FPS on a single consumer RTX 4090 GPU. This makes it practical not only as a visual prediction model, but as an interactive engine for collecting policy-training data and evaluating robot policies reproducibly. It has multiple applications.

Application 1: scalable data generation

If our simulator is good enough, can demonstrations collected using this world model actually replace real demonstrations for training imitation learning policies?

To answer this, we collected demonstrations entirely inside our world simulator and then trained imitation learning policies on 0% real-world data and 100% generated data. We deployed the trained policies directly on the real robot. The deployed policies do not just complete the tasks. They remain robust under continuous human perturbations. This demonstrates that the generated data quality is comparable to real data.

Application 2: faithful policy evaluation

Evaluating a robot policy in the real world is hard: you need to reset the scene, rerun the policy many times, and compare across checkpoints under matched conditions. In practice, this is not scalable and reproducible.

In contrast, we could roll out policies in our world model for reproducible evaluation. We evaluated the same four policies (DP, ACT, π0, π0.5) inside the simulator and on the real robot, with the same initial configurations. Each point in our evaluation is a policy checkpoint, scored in both real world and world model. We observe a strong correlation between the two scores, across tasks and across policies. Qualitatively, good policies that succeed in the simulator also succeed on the real robot, and bad policies that fail in the simulator also fail on the real robot.

Looking forward

Action-conditioned video prediction has long shown great potential for robotics, but existing approaches have been either too slow for interactive use or too brittle under long-horizon inference. By addressing both of these — stability and computational efficiency — the Interactive World Simulator becomes a practical engine for two things at once: training imitation policies on simulator-generated data that perform comparably to those trained on real-world demonstrations, and evaluating policies in a way that closely tracks real-world performance.

Going forward, we will extend the framework to more diverse environments and increasingly complex manipulation tasks, to further unlock the potential of large-scale robotic data. An important direction for future work is to study how the performance of world models scales with increasing amounts of interaction data and computational resources. Understanding these scaling behaviors could guide the design of larger and more capable world models for robotics.

By Holly Hartigan

A team of Cornell undergraduates beat 95 other teams to take the grand prize at The Farm Robotics Challenge with their invention: an autonomous robot that kills weeds with electricity.

Their robot can travel through a vineyard or orchard without a human operator, zapping weeds with a small amount of electricity, saving labor and energy and preventing crop loss, without the use of herbicides.

Led by Andrew James, an agricultural sciences major in the College of Agriculture and Life Sciences (CALS), the team of agricultural specialists and engineers studied the existing electrical weeding technology, developed their own low-energy system and built a working prototype over the course of four intense months.

Natalia Kurz, a biological engineering major in CALS, said the project required a lot of late nights. “There were fears for us, like, was it just going to be for nothing?”

Now, James and his co-founders are using the $50,000 grand prize to form a company – Rootline Robotics – to continue working on the robot. Agricultural technology firm Reservoir sponsored the award and will host the startup at its incubator in Sonoma, California.

“I’ve always been interested in building a startup within the ag-robotics space,” James said. “So after winning this competition and seeing all the amazing support from so many different industry stakeholders in this really exciting collaboration, it makes a lot of sense to keep going.”

The problem of weeds

Weeds are a huge challenge in orchards and vineyards because they steal water and nutrients, especially in the spring, according to Steve Selin, owner of South Hill Cider in Ithaca.

“The weed pressure is very strong, and the grasses grow right up to the trees because we can’t afford to weed whack or mulch them as much as we would like to in order to control them,” he said. “If the trees get really stressed out, they’ll just drop the fruit.”

To control weeds, organic growers typically employ string trimming, mowing and mulching, which are all very labor intensive. Existing electric weeders require an operator, consume a lot of energy and cost $150,000, on average, which is out of reach for most growers, James said.

Selin provided feedback to the students as they developed their robot and said he is excited to try out Rootline’s new technology.

“In May and June, if you could do something to knock the weeds back enough that they’re not going to compete with the trees, then the rest of the year you wouldn’t have to worry about it,” he said. “You can let them come back and have positive impacts, like shading the soil, which would help the soil microbes to have a healthier ecosystem.”

The interdisciplinary nature of robots

“Robotics is many different systems in one,” Kurz said. “It is interdisciplinary by definition.”

Next, the Rootline team will work on improving its technology and validating it with growers before bringing it to market, all while Kurz and Neiss finish their degrees.

Neiss said the way the team came together felt a bit like destiny.

“If you have many different minds in one room that gives you the ability to reach solutions that are going to be the most feasible and effective,” he said.

This year, RoboCup took place in Incheon, South Korea, from 2-6 July. The event saw teams take part in competitions, training sessions, and a symposium. Take a look at what the participants got up to in our round up from social media.

The wait is over!

RoboCup 2026 officially begins today!

Join us for an exciting first day!

Incheon Convention Center (Songdo Convensia)

For the full schedule, visit:
https://t.co/T5mjdVMhH3#RoboCup2026 #RoboCup #Robotics #AI pic.twitter.com/wHZkfM69jL

— RoboCup Federation (@robocup_org) July 1, 2026

View this post on Instagram

A post shared by RoboCup Federation (@robocup.official)

View this post on Instagram

A post shared by BerlinUnited (@berlin_united_robocup)

Can't make it to Incheon? We've got you covered!

Starting July 2, follow all the excitement of RoboCup 2026 live from anywhere in the world!

Where to watch:
YouTube: @RoboCup
Twitch: robocupofficial#RoboCup2026 #RoboCup #LiveStreaming #Robotics #AI #Incheon pic.twitter.com/p3wm8mr3CV

— RoboCup Federation (@robocup_org) July 1, 2026

View this post on Instagram

A post shared by RoboCupJunior Egypt (@rcjegypt)

Little Booster K1, trying to get up after falling mid football game!
#robocup2026

[image or embed]

— Raghav Arora (@ra-aurora.bsky.social) 6 July 2026 at 05:05

View this post on Instagram

A post shared by Colegio Leonardo Da Vinci (@davinci_colegio)

View this post on Instagram

A post shared by Tech United (@techunitedeindhoven)

Big win at #RoboCup2026!

4 Mini Pi plus teams reached the Top 8, with Hamburg Bit-Bots taking .

Open platforms accelerate robotics. #HumanoidRobots #Robotics #OpenSource #EmbodiedAI #ROS2 #Automation #PhysicalAI pic.twitter.com/BAZeQGHuNp

— High Torque Robotics (@HTrobotics) July 5, 2026

View this post on Instagram

A post shared by B-Human (@bhumnbhumn)

View this post on Instagram

A post shared by Fernanda Bufon Färber (@fernandabufon)

Straight from RoboCup 2026, at Incheon, Korea.
We added an external lidar, and a literal backpack to the booster T1 to achieve accurate indoor localisation and navigation.
@texasrobotics.bsky.social #Robocup2026

[image or embed]

— Raghav Arora (@ra-aurora.bsky.social) 6 July 2026 at 05:01

View this post on Instagram

A post shared by Robot Club Toulon (@robotclubtoulon)

A team from China's Tsinghua University successfully defended its Humanoid League title at RoboCup 2026 on Sunday in South Korea. #RoboCup2026 #HumanoidLeague #TsinghuaUniversity #robotsoccer pic.twitter.com/88gKYpxL3J

— China Xinhua News (@XHNews) July 6, 2026

View this post on Instagram

A post shared by autonOhm (@autonohm_official)

This is RoboCup 2026!

Nearly 3,000 participants from 40 countries, united by one passion: advancing robotics and artificial intelligence!
Thank you to everyone who is part of this incredible journey! #RoboCup2026 #Robotics #AI #Incheon pic.twitter.com/X5enZLypmF

— RoboCup Federation (@robocup_org) July 4, 2026

The host venue for ICRA 2026: Messe Wien, also known as VIECON.

Workshop on robot ethics

My week at ICRA began with the 2nd ICRA 2026 Workshop on Robot Ethics: Ethical, Legal and User Perspectives in Robotics & Automation (WOROBET). WOROBET provided a space for researchers to share ideas, thoughts, and concerns on the future of robot-human interaction, and how to create ethical frameworks to navigate this rapidly changing technology. 

Yasuhisa Hirata, Professor at Tohoku University, began by presenting his vision of a world with physically assistive robots, such as detachable exoskeletons or cycling wheelchairs. These tools can affect people’s sense of self-efficacy and motivation, and there are a host of ethical implications that come with this – how do you help people just enough to build their confidence, without slipping into deception? 

We then heard from Prof. Minoru Asada from Osaka University, who discussed his aim to implement pain signals into robots, so they can experience the world as we do. This was a highly interesting and niche proposition that brought up more ethical questions than we have answers for at the moment. Perhaps most pertinent to a technical conference: is a sense of embodied morality necessary for true intelligence? 

Alan Winfield, Professor of Robot Ethics at UWE Bristol, presented a vision of robotics that acted as a counterweight to Prof. Asada’s: robots as tools, not potential beings. This also somewhat reflects differing attitudes in Eastern vs Western cultures. Thinking about the practical issues we are likely to face in the near term, he outlined a framework for social robot accident investigation. In his view, robot ethics is not just an engineering problem, but requires appropriate social and governance frameworks, just as we do for aviation. His talk also emphasised the risk that programming ethics into robots runs the risk of removing moral responsibility from the roboticist, and can always give rise to unethical robots via malicious hacking. The theme that we must focus on human morality, as opposed to machine morality, was repeated throughout the day. 

After a morning of differing ideas and visions of what robot-human interaction could be, we were invited to ground this into a real robot social care scenario by Praminda Caleb-Solly, Professor of Embodied Intelligence at the University of Nottingham. In this red-teaming exercise, we examined the safety risks and possible mitigations of an assistive robot for a schoolteacher recovering from a stroke at home. Our group discussion circled around human agency: how ethical is it to make design choices for people that take away some of their autonomy, in the name of their best interests? As robots in social care will become a more urgent need in the years to come, these questions may become more salient. 

I left WOROBET with plenty to think about, and a renewed sense of appreciation for the ethicists who are already grappling with the problems that are to come. I hope that progress in robot ethics keeps pace with progress in robotics, so that we are well prepared as robots become more of a part of our daily lives.

Welcome to the jungle – the robot exhibition floor

Every time you entered the exhibition hall, you were greeted by one of the child-sized Booster robots, either playing football, dancing, or demonstrating some kung fu. They were always an endearing welcome to the sea of robots. 

The robots ranged from the endearing to the uncanny, but the common thread was their technical capabilities were astounding. Veteran attendees consistently remarked on how much the robots had improved year on year.

The cute.

The slightly uncanny.

The below clips show the robots that most caught my eye. After admiring a phosphorescent, Stranger Things-esque robotic flower display, I was blown away by the D1-modular robot from Direct Drive. Unlike most robot dogs, it can split into two halves, with the ability to jump, twist, and traverse difficult terrain. 

Sharpa’s North was always a friendly face, waving and making love-hearts at visitors to its booth. Around the back of the booth, you could even challenge it to a round at blackjack. We saw humanoids zipping up rucksacks and trying to fold laundry. Enchanted Tools’ social care robot, Mirokaï, was an unusual sight among the mass of black and steel, with a bright orange body, feline ears and an orange, furry face. Tesollo’s humanoid spent much of its time at ICRA using its long, wavering arms to pick up fruit and drop it into baskets, with impressive dexterity. Vietnamese company Vinrobotics’ humanoid offering was reminiscent of the Cybermen, with its gently wheezing joints, but the team assured me it was much friendlier. The pint-sized Boosters were almost always playing football, not far from their similarly sized Agibot cousins. 

However, humanoids didn’t steal the show this year, as they have done previously. The big trend this year was robotic hands, and the levels of dexterity were truly impressive. Closing the gap between human and robot abilities here would unlock whole new swathes of tasks to automate – and who wouldn’t want a robot folding their laundry? 

Industrial challenges: solving dexterity 

Tackling the dexterity challenge really defined the industrial talks for me this year. A talk by ARIA’s program director, Prof. Jenny Read, demonstrated how the UK government is already laying the groundwork here.

They outlined their funding proposals as part of their Smarter Robot Bodies program, which is split into two branches: robot locomotion and robot dexterity. The robot locomotion branch aims to enable robots to traverse messy, unpredictable physical environments, while the robot dexterity branch will try to break the bottleneck of adept physical manipulation by robotic hands. With the programme set to launch in early 2027, it’ll be exciting to see what kind of innovations this attracts. 

One standout innovation in the realm of dexterity was TARS, a record-setting newcomer in the Chinese robotics market. Co-founded by Dr Ding Wenchao just 18 months ago, TARS has already raised the most funding in angel and pre-seed rounds of any company in the Chinese embodied intelligence sector, and achieved a Guinness World Record for robotic flexible wiring-harness insertion completed in one hour. 

TARS’ DexHand. Image credits: TARS.

DexHand is a 1:1 model of the human hand, even replicating the 21 degrees of freedom we have in the wrist joint and hand. According to TARS, “it can interpret tactile data to distinguish slipperiness, roughness, and hardness in real time and perform 26 English alphabet hand gestures with high-precision finger control.” I was given the chance to control the hand using my own, and was impressed by how well it emulated my own movements. I was also impressed by their humanoid zipping up a backpack – despite the technical abilities of all the robots on the floor, few were able to perform tasks which required such fine motor skills. 

TARS robot zipping up a rucksack.

On Thursday, Dr Wenchao Ding delivered an industry keynote where he presented TARS’ roadmap, charting the path from academia to industrial deployment. With an impressive academic team behind them, TARS may be one to watch. 

Plenary talks

Aside from the wealth of invention and innovation taking place in the exhibition hall, the breadth and depth of academic research at ICRA was fantastic. The plenaries especially gave an insight into the research trends that are currently defining the field. 

Ken Goldberg delivered an electrifying plenary, titled “A Tale of Two Cultures:  Can Agentic Coding Close the Gap?” In this talk, he called for a step change to close the data gap faced by robot manipulation. With the rise of diffusion models and LLMs, it is clear that big data has solved computer vision and language. He challenged the audience – when will the ChatGPT moment for robotics come? With state spaces larger than 50 dimensions in robotics, there is not enough training data to close this gap. Currently, the data required to train vision-language models is equivalent to 100,000 years of real physical experience. 

According to Goldberg, the two dominant cultures in engineering – model free “good old fashioned” engineering (GOFE), and, the currently more popular model based engineering.  GOFE encapsulates rigorous engineering methods that pre-date AI, but may have been slightly forgotten about in the AI wave of recent years. He highlighted how in his career he has always been working to bridge the gap between two cultures: from science and art to robotics and automation. 

He outlined four possible solutions to the data gap:

1. Simulations – these work incredibly well for locomotion and body control, but less so for manipulation due to the number of forces and instabilities involved.
2. World models – they do not currently properly capture the physics, and hallucinations can be problematic. 
3. Human teleoperation – this is currently big business and a good way to obtain high quality data. However, the largest dataset is currently only equivalent to a year’s worth of data. 
4. Real data from functioning robots – this is less commonly used, but can be powerful. 

Prof Goldberg described how he used the fourth approach in his robotic delivery packing company, Ambi Robotics, for 22 years. Picking up bags is an example of variational automation –  one task is done repeatedly, but with different initial conditions each time. From this rich dataset, they created a generative model to train robots in the best way to pick up bags. Here, they close the gap between model-free and model-based methods – Ambi exploits both to achieve industry-leading results. This plenary served as a call for other researchers to use their own production data to do the same. 

During Thursday’s keynote on Robot Learning, Planning & Foundation Models, Stefanie Tellex from Brown University gave a compelling talk titled “Towards Complex Language in Partially Observed Environments”. While current research is bounded in known, predictable scenarios, using action-based language, this does not reflect what the real world is like, nor how people would naturally communicate with robots. Prof. Tellex described her work creating robots that can understand complex, goal-based commands in only  partially observed, dynamic environments, and outlined the grounded Turing test – a reimagining of the Turing test for embodied AI. 

An example of a robot performing a goal-based task in a dynamic environment. Credits: Tellex et al, 2026.

Both plenaries spoke to current pinch points in robotics: data, reasoning, and operating in complex real-world environments. It’ll be interesting to see what solutions are developed in the coming years. 

Science communications crash course

One of my favourite parts of ICRA was delivering the Science Communications Crash Course. Along with Robohub Executive Trustee Sabine Hauert, IEEE Spectrum Senior Editor Evan Ackerman, and IEEE Spectrum Community Manager Kohava Mendelsohn, we gave our guidance on effective science communication to an audience of 100 academics. It was encouraging to see so many people interested in communicating their research effectively – it is a crucial skill, especially in the era of AI and robotics when mainstream narratives can be hijacked by doom-mongering, hype, and corporate interests. More academics communicating their work clearly and neutrally will go a long way to grounding our societal discussion in technical reality, not sci-fi futures. 

Sabine kicking off the science communications crash course. Image credits: Taraja Arnold

Delivering my part of the course. Image credits: Taraja Arnold

Art and robotics 

The arts and robotics section was rich and interesting. There was a lot to visually take in, with constant background music from a robotic saxophone. 

Masatoshi Hamanaka’s robotic saxophone. Image credits: ©Denes Erdos – Your Event Photographer

PET – marked by a large “PET ME” sign – was a white and orange mass of connecting, rotating pyramids, that gently pulsed and hummed in response to touch, responding by curling towards or away from you depending on how you touched it, The effect was strangely lifelike. 

Rhombus Research presented “Reptile: A Bio-Mimetic Choreography Engine for V2X and A2X Swarms”. This artistic simulation visualises the contracts negotiated within autonomous vehicle and swarm fleets in dreamy blue, browser-based visualisation. Performance artist and former Cirque de Soleil acrobat Silke Grabinger explored human-robot interaction via her piece, AREYOUARE. 

Silke Grabinger performing AREYOUARE. Image credits: ©Denes Erdos – Your Event Photographer

I’d also like to acknowledge some of the video creators I met there. YouTubers Back to Engineering and the. Amazing, PhD are making some fantastic videos about physical AI. 

Nothing lasts forever – ending with the robot parade 

ICRA 2026 ended with the robot parade, which attracted quite a crowd. See if you can spot the panda, dragon, and headless humanoid!

Seeing all the robots gathered together was a real spectacle. The technology on display was state-of-the-art, and it only improves year on year. What struck me most was the sense that robotics is moving from proving what is possible to tackling the remaining barriers to real-world deployment. Across the exhibition floor, industry keynotes, and plenary talks, the focus was often on the same challenges: dexterity, data, and operating reliably in complex environments. Solving these bottlenecks will open up new avenues to real-world applications of robots. 

While workshops such as WOROBET highlighted important questions around ethics, agency, and governance, the overwhelming emphasis at ICRA 2026 was on capability. Researchers and companies alike are working to close the gap between what robots can do in carefully controlled demonstrations and what they can do in the messy reality of the world outside the lab. Judging by the pace of progress on display in Vienna, that gap may be narrowing faster than many of us expected. I hope that the kinds of conversations around human-robot interaction and ethics that were commonplace at WOROBET and in the Arts exhibition space will become more mainstream – we may need to face the questions that they raise sooner than we think. 

Note: Where image and video credits are not stated, they belong to Ella Scallan.

RoboCup 2026 kicked off today in Incheon, South Korea, with the league competitions running until 5 July. It’s an exciting time for RoboCup, as there have been some updates to the leagues and competition format. Most prominently, the soccer leagues will have a primary focus on humanoid robots. In a series of daily updates, we’ll be bringing you the latest results, videos, and news from the humanoid soccer league.

This year, the humanoid league is split into three sizes: large division, middle division, and small division. There are 18 teams participating the small division, 16 in the middle, and an impressive 22 competing in the large division.

The first two days of competition are be devoted to the seeding round. Teams play using the Swiss-system of ranking to decide who gets through to the knockout stages.

Livestreams from the different fields of play can be found here.

At the end of the first day of competition, teams in the small and middle divisions have played two games each, with the large division ending the day part way through the second round. Early leaders in the small division, with the full six points from two games are: GeoHBots, CAU Mountain&Sea, and Hamburg Bit-Bots. There are also three teams in the middle division who have claimed the six point haul: B-Human, RoboRoos, and HTWK Robots.

You can watch a short summary of the first day, including some of the robots in action, from KBS News:

This short video from RoboCup gives a flavour of the day’s happenings, which also featured the opening ceremony.

View this post on Instagram

A post shared by RoboCup Federation (@robocup.official)

We will be back tomorrow with further updates on competition results, some highlights from the day, and insights into how the event is progressing.

Useful links

• RoboCup 2026 homepage
• Humanoid league YouTube channel
• Humanoid league results page
• RoboCupSoccer homepage

By Adam Zewe

In the aftermath of a devastating earthquake, unpiloted aerial vehicles (UAVs) could fly through a collapsed building to map the scene, giving rescuers information they need to quickly reach survivors. 

But this remains an extremely challenging problem for an autonomous robot, which would need to swiftly adjust its trajectory to avoid sudden obstacles while staying on course.

Researchers from MIT and the University of Pennsylvania developed a new trajectory-planning system that tackles both challenges at once. Their technique enables a UAV to react to obstacles in milliseconds while staying on a smooth flight path that minimizes travel time. 

Their system uses a new mathematical formulation that ensures the robot travels safely to its destination along a feasible path, and that is less computationally intensive than other techniques. In this way, it generates smoother trajectories faster than state-of-the-art methods.

The trajectory planner is also efficient enough for real-time flight using only the robot’s onboard computer and sensors. 

Named MIGHTY, the open-source system does not require proprietary software packages that can cost hundreds of thousands of dollars. It could be more readily deployed in a wider variety of real-world settings.

In addition to search-and-rescue, MIGHTY could be utilized in applications like last-mile delivery in urban spaces, where UAVs need to avoid buildings, wires, and people, or in industrial inspection of complex structures, such as wind turbines.

“MIGHTY achieves comparable or better performance using only open-source tools, which means any researcher, student, or company — anywhere in the world — can use it freely. By removing this cost barrier, MIGHTY helps democratize high-performance trajectory planning and opens the door for a much broader community to build on this work,” says Kota Kondo, an aeronautics and astronautics graduate student and lead author of a paper on this trajectory planner.

Kondo is joined on the paper by Yuwei Wu, a graduate student at the University of Pennsylvania; Vijay Kumar, a professor at UPenn; and senior author Jonathan P. How, a Ford professor of aeronautics and astronautics and a principal investigator in the Laboratory for Information and Decision Systems (LIDS) and the Aerospace Controls Laboratory (ACL) at MIT. The research appears in IEEE Robotics and Automation Letters.

Overcoming trade-offs           

When Kondo was a child, the Fukushima Daiichi nuclear accident occurred following the Great East Japan Earthquake. With school cancelled, Kondo was stuck at home and watched the news every day as workers explored and secured the reactor site. Some workers still had to enter hazardous areas to contain the damage and assess the situation, exposing them to high doses of radioactive material.

“I became passionate about creating autonomous robots that can go into these dynamic and dangerous situations, then come back and report to humans who stay out of harm’s way,” Kondo says.

This task requires a strong trajectory planner, which is software that decides the path a robot should follow to safely get from point A to point B. 

But many existing systems force tradeoffs that limit performance. 

While some commercial systems can rapidly generate smooth trajectories, they can cost hundreds of thousands of dollars. Open-source alternatives often underperform compared to commercial solvers or are difficult to use.      

With MIGHTY, Kondo and his colleagues developed an open-source system that produces high-quality, smooth trajectories while reacting to obstacles in real-time, and which runs fast enough for flight using only onboard components.

To do this, they overcame a key challenge that limits many open-source systems. 

These methods usually estimate how long it will take the robot to get from point A to point B as a first step. From that fixed estimation of travel time, the planner finds the best path to reach the destination.

While using a fixed travel time allows the planner to rapidly generate a trajectory, it has drawbacks. For one, if the UAV must go far out of its way to avoid obstacles, it could be forced to crank up the speed to meet the fixed travel-time budget. This makes it harder to avoid sudden hazards.

A MIGHTY method

Instead, MIGHTY uses a mathematical technique, called a Hermite spline, that optimizes the travel time and flight path together, in a single step, to form a smooth trajectory that can be precisely controlled.

“Optimizing the spatial and temporal components together gets us better results, but now the optimization becomes so much bigger that it is harder to solve in a feasible amount of time,” Kondo says.

The researchers used a clever technique to reduce this computational overhead. 

Instead of generating a trajectory from scratch each time, MIGHTY makes an initial guess of a trajectory. Then it refines the trajectory through an iterative optimization, using a map of the scene generated by the UAV’s lidar sensors.

“We can make a decent guess of what the trajectory should be, which is a lot faster than generating the entire thing from nothing,” Kondo says.

This enables MIGHTY to react in real-time to unknown obstacles while keeping the trajectory smooth and minimizing travel time. The system utilizes the UAV’s onboard components, which is important for applications where a robot might travel far from a base station.

In simulated experiments, MIGHTY needed only about 90 percent of the computation time required by state-of-the-art methods, while safely reaching its destination about 15 percent faster than these approaches. 

When they tested the system on real robots, it reached a speed of 6.7 meters per second while avoiding every obstacle that appeared in its path.

“With MIGHTY, everything is integrated in one piece. It doesn’t need to talk to any other piece of software to get a solution. This helps us be even faster than some of the commercial solvers,” Kondo says.

In the future, the researchers want to enhance MIGHTY so it can be used to control multiple robots at once and conduct more flight experiments in challenging environments. They hope to continue improving the open-source system based on user feedback.

“MIGHTY makes an important contribution to agile robot navigation by revisiting the trajectory representation itself. Hermite splines have already been successfully used in visual simultaneous localization and mapping, and it is nice to see their advantages now being exploited for trajectory planning in mobile robots. By enabling joint optimization of path geometry, timing, velocity, and acceleration while retaining local control of the trajectory, MIGHTY gives robots more freedom to compute fast, dynamically feasible motions in cluttered environments,” says Davide Scaramuzza, professor and director of the Robotics and Perception Group at the University of Zurich, who was not involved with this research.

This research was funded, in part, by the United States Army Research Laboratory and the Defense Science and Technology Agency in Singapore.

Claire chatted to Edward Mehr from Machina Labs about their RoboCraftsman that shapes complex metal parts for the aerospace, defence, and automotive industries.

Edward Mehr is an entrepreneur and engineer specializing in advanced manufacturing, robotics, and artificial intelligence. As the Co-Founder and CEO of Machina Labs, he leads efforts to integrate AI-driven robotics into flexible, on-demand production systems. Under his leadership, Machina Labs is reshaping how industries such as aerospace, defence, and automotive approach metal forming and modern manufacturing. Before founding Machina Labs, Ed worked at leading technology companies, including Relativity Space, Averon, SpaceX, Google, and Microsoft.

By Kartikeya Walia, Nottingham Trent University

A table tennis robot has outperformed elite players in recent evaluations. The robot, called Ace, marks a significant step toward artificial intelligence (AI) systems that can operate in fast, uncertain, real-world environments.

In the tests, the autonomous robot won three out of five matches against elite players – competitive athletes with over ten years’ experience and an average of 20 hours weekly training. The robot, developed by Sony AI, lost both matches against players in professional Japanese leagues, but did win a game against one of them. The system is described in detail in a recent paper published in Nature.

AI has spent decades mastering games. It has repeatedly outperformed the best humans in everything from complex video games like StarCraft II to chess – where modern programs now far exceed human ratings.

Landmark systems such as Deep Blue and AlphaGo have confirmed that, given clear rules and enough data, AI can achieve superhuman performance. But these victories all shared one key feature: they happened in controlled, digital environments.

At first glance, table tennis might seem like an unusual benchmark for artificial intelligence. In reality, it is one of the most demanding imaginable. The ball can travel faster than 20 metres per second, giving players less than half a second to react.

On top of that, spin introduces enormous complexity. A ball rotating at extreme speeds can curve mid-air and rebound unpredictably off the table. For humans, interpreting spin is largely intuitive. For robots, it has been a longstanding obstacle.

This robot can beat you at table tennis (Nature).

Earlier table tennis robotic systems such as Forpheus, developed by Japanese company Omron, addressed this by simplifying the game – using controlled ball launchers, limiting movement, or ignoring spin altogether. More recent iterations have aimed for interaction, but still operate under constrained conditions.

Ace does none of this. It plays with standard equipment, on a regulation table,
against human opponents who are free to use the full range of shots.

How Ace works

Ace’s performance relies on three key innovations: how it sees the world, how it
decides what to do, and how it carries out those actions. First, let’s deal with how Ace sees the world. Traditional cameras struggle with fast motion, often producing a blur or missing critical details.

Ace instead uses three “event-based” vision sensors, which detect changes in light rather than capturing full images at fixed intervals. These are complemented by nine high-speed cameras that track the environment, including the opponent and their racket.

Together, these systems enable high-speed gaze control (the technology that enables a robot to direct its sensors to focus on specific things) and allow the robot to follow the ball with exceptional real-time precision.

By tracking markings on the ball, where professional players can generate spin approaching 9,000 revolutions per minute (rpm), the system can estimate spin in real time, something that has long challenged robotic systems.

How Ace’s gaze control system works (Sony AI and Nature).

The second important innovation is how Ace decides what to do. Knowing where the ball is going is only half the problem; the robot must also respond instantly. Ace uses deep reinforcement learning, trained in simulation over millions of virtual rallies, including self-play.

It continuously generates movement commands for its multi-jointed robotic arm, recalculating trajectories every few tens of milliseconds while avoiding collisions with the table or itself.

The third innovation is how Ace carries out its actions. To match the speed of human elite players, the robot is built around a high-performance arm combining two prismatic (sliding) and six revolute (rotational) joints. This enables rapid sideways motion and precise striking. There is both a table tennis racket and a mechanism for ball handling, allowing one-armed serves.

Crucially, the system is engineered for high-speed interaction: lightweight structures and optimised actuation (the mechanisms in a robot that convert energy into mechanical force) allow Ace to return balls at speeds approaching 20 metres per second. This enables sustained, competitive rallies with skilled human players.

Ace makes a split section change when the ball hits the net (Sony AI and Nature).

What makes this particularly notable is the transition from simulation to reality. Many AI systems perform well in virtual environments but fail when exposed to real-world noise and uncertainty. Ace demonstrates that this “sim-to-real” gap can be meaningfully reduced.

One moment during a rally with an elite player illustrates the way that Ace has leapt over this gap. When a predicted ball trajectory suddenly changed after clipping the net, Ace reacted almost instantly, returning the shot and winning the point. What makes Ace particularly significant is therefore not just its performance, but its ability to operate reliably under real-world uncertainty.

Why this matters beyond sport

A robot returning high-speed topspin shots may be entertaining, but the implications go far beyond table tennis. In manufacturing, for example, robots are typically confined to highly structured tasks.

The real challenge is adaptability, handling irregular objects, responding to variation. This is particularly relevant for next-generation robots operating in unstructured environments.

To function effectively in homes, hospitals or construction sites, robots must be able to predict, adapt and respond to constantly changing conditions. The same predictive and control capabilities that allow Ace to respond to unpredictable shots could enable more flexible, responsive automation.

Most industrial robots are kept behind safety barriers because they cannot respond to unexpected human behaviour. Zhu Difeng

There are also implications for human–robot interaction. Most industrial robots are kept behind safety barriers because they cannot react quickly or reliably enough to unexpected human behaviour. Ace operates at the edge of human reaction time, suggesting a future where robots can safely collaborate with people in shared spaces.

More broadly, this work represents a shift in what AI is expected to do. The next frontier is not just intelligence in abstract problem-solving, but intelligence embedded in the physical world. The gap between simulations and reality needs filling, and this is a big step forward.

What humans still do better

Professional players were still able to exploit Ace’s limitations – particularly in reach, speed, and the ability to handle extreme or highly deceptive shots. This highlights that intelligence is not just about prediction and control, but also about physical embodiment. Humans combine perception, movement and strategy in ways that remain difficult to replicate.

Interestingly, systems like Ace may end up enhancing human performance rather
than replacing it. As one former Olympic player observed after facing the robot,
seeing it return seemingly impossible shots suggests humans might be capable of more than previously thought.

Kartikeya Walia, Senior Lecturer, Department of Engineering, Nottingham Trent University

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

By David Nutt

The way bugs and birds flap their wings may look effortless, but the dynamics that keep them aloft are dizzyingly complex and difficult to quantify.

Cornell researchers created a computational model that shows the effect of insects’ morphology on stabilizing their flight. The findings could lead to a new way to understand the evolution of animal flight while also providing a blueprint for designing flapping-wing robots.

The study published May 1 in Proceedings of the National Academy of Sciences. The research was led by Z. Jane Wang, professor of physics and mechanical and aerospace engineering in the College of Arts and Sciences and Cornell Duffield College of Engineering, respectively.

The effort began more than a decade ago, when Wang set out to understand how the neural circuitry in fruit flies evolved to control flight stability. By creating a 3D computational simulation, Wang’s team showed that fruit flies sense the orientation of their bodies every time they beat their wings, about one beat every 4 milliseconds, in order to stabilize themselves.

However, in order to study flight stability in all insects, the researchers would need to build an efficient computational tool to simulate a huge number of species.

“Previous studies, including ours, have always started with models of real insects, so we’re limited by the things we observe,” Wang said. “We miss all the other configurations that are also possible for flight.”

Wang and Owen Wetherbee, the new paper’s first author, distilled the 3D model into a new version that retained the key physics of the body-wing coupling and unsteady aerodynamics. The resulting equations revealed the critical physical parameters: wing to body mass ratio, wing loading, wing hinge position, wing beat frequency and wing motion amplitude. Taken together, they form what Wang calls a “five-dimensional morphological and kinematic space.”

“The power of this model is to give us something much more explicit than what we had before,” she said. “We knew the fundamental physics. By capturing the essential physics in the new model, we can understand each piece conceptually as well as facilitate computation to explore a large parameter space.”

The analyses of the computational results in 5D resulted in two explicit formula that provide a succinct metric for stability. These criteria capture the subtle and often ignored coupling between wing inertia and the body, which depends on the interplay among wing flap frequency, hinge placement, and wing and body mass ratios in order to achieve a kind of anti-resonance state. This sweet spot allows the flapping winged animal to control its body oscillations and remain aloft – a state known as passively stable flight – despite air perturbations that would normally cause it to tumble.

“All of a sudden, we found that many forms of flapping flight have passive stability, which surprised us initially, because works so far showed that most insects, except one or two, are passively unstable, hence the necessity for neural circuitry to control them,” Wang said. “But when we expanded the morphological space, we realized that what we studied before are but a few dots in this new view.”

Now that the researchers can characterize the stability boundary, they can offer a concrete design principle for realizing stable flapping flight in robots – something that has stumped roboticists for decades.

“In principle, this offers a completely new route for designing a robotic flapping-winged machine,” Wang said. “Instead of relying on extensive feedback control, which is only partially successful, our results suggest that we can tune the shape and the frequency of the flapping devices such that, according to these two rules, we may find the flyers are passively stable already. This would greatly simplify flight control.”

The new model allows this design work to be done with faster and simpler computation, and the ability to model stability traits also points to a new way for classifying winged animals and charting their evolution.

“During evolution, various traits are selected, but we don’t have much idea about what they are, let alone understand why they are being selected and how they evolve, apart from a very few examples, such as an eye,” Wang said. “This project brings new quantitative methods to study these very big questions in both biology and robotics. Mathematical modeling allows us to go beyond our own ideas and preconceptions to tackle these large questions.”

The research was supported by the National Science Foundation.

By Adam Zewe

Robotically assembled building blocks could be a more environmentally friendly method for erecting large-scale structures than some existing construction techniques, according to a new study by MIT researchers.

The team conducted a feasibility study to evaluate the efficiency of constructing a simple building using “voxels,” which are modular 3D subunits that assemble into complex, durable structures.

After studying the performance of multiple voxels, the researchers developed three new designs intended to streamline building construction. They also produced a robotic assembler and a user-friendly interface for generating voxel-based building layouts and feeding instructions to the robots.

Their results indicate this voxel-based robotic assembly system could reduce embodied carbon — all of the carbon emitted during the lifecycle of building materials — by as much as 82 percent, compared with popular techniques like 3D concrete printing, precast modular concrete, and steel framing. The system would also be competitive in terms of cost and construction time. However, the choice of materials used to manufacture the voxels does play a major role in their carbon footprint and cost.

While scalability, durability, long-term robustness, and important considerations like fire resistance remain to be explored before such a system could be widely deployed, the researchers say these initial results highlight the potential of this approach for automated, on-site construction.

“I’m particularly excited about how the robotic assembly of discrete lattices can enable a practical way to apply digital fabrication to the built environment in a way that can let us build much more efficiently and sustainably,” says Miana Smith, a graduate student in the Center for Bits and Atoms (CBA) at MIT and lead author the study.

She is joined on the paper by Paul Richard, a graduate student at École Polytechnique Fédérale de Lausanne in Switzerland and former visiting researcher at MIT; Alfonso Parra Rubio, a CBA graduate student; and senior author Neil Gershenfeld, an MIT professor and the director of the CBA. The research appears in Automation in Construction.

Designing better building blocks

Over the past several years, researchers in the Center for Bits and Atoms have been developing voxels, which are lattice-structured building blocks that can be assembled into objects with high strength and stiffness, like airplane wings, wind turbine blades, and space structures.

“Here, we are taking aerospace principles and applying them to buildings. Why don’t we make buildings as efficiently as we make airplanes?” Gershenfeld says, based on prior work his lab has done on voxel assembly with NASA, Airbus, and Boeing.

To explore the feasibility of voxel-based assembly strategies for buildings, the researchers first evaluated the mechanical performance and sustainability of eight existing voxel designs, including a cuboctahedron made from glass-reinforced nylon and a Kelvin lattice made from steel.

Based on those evaluations, they developed a set of three voxels using a new geometry that could be more easily assembled robotically into a larger structure. The new design, based on a high-strength and high-stiffness octet lattice, mechanically self-aligns into rigid structures.

“The interlocking nature of these voxels means we can get nice mechanical properties without needing to have a lot of connectors in the system, so the construction process can run a lot faster,” Smith says.

To accelerate construction, they designed a robotic assembly system based on inchworm-like robots that crawl across a voxel structure by anchoring and extending their bodies. These Modular Inchworm Lattice Assembler robots, or MILAbots, use grippers on each end to place voxel building blocks and engage the snap-fit connections.

“The robots can assemble the voxels by dropping them into place and then stepping on them to have the pieces interlock. We can do precise maneuvers based on the mechanical relationship between the robots and the voxels,” Smith explains.

The team studied the embodied carbon needed to fabricate their new voxel designs using three materials: plastic, plywood, and steel. Then they evaluated the throughput and cost of using the robotic assembly system to build a simple, one-story building. The researchers compared these estimates with the performance of other construction methods.

The MILAbot’s unique legs, seen here in close up. “The robots can assemble the voxels by dropping them into place and then stepping on them to have the pieces interlock,” Miana Smith explains. Credit: Courtesy of the researchers.

Potential environmental benefits

They found that most existing voxels, and especially those made from plastics, performed poorly compared to existing methods in terms of sustainability, but the steel and wood voxels they designed offered significant environmental benefits.

For instance, utilizing their steel voxels would generate only 36 percent of the embodied carbon required for 3D concrete printing and 52 percent of the embodied carbon of precast concrete. The plywood voxels had the lowest carbon footprint, requiring about 17 percent and 24 percent of the embodied carbon needed, respectively.

“There is still a potential viable option for a plastics-based voxel approach, we just have to be a bit more strategic about which types of plastics, infills, and geometries we use,” Smith says.

In addition, projected on-site assembly time for the steel and wood voxel approaches averaged 99 hours, whereas existing construction methods averaged 155 hours.

These speed benefits rely on the distributed nature of voxel-based assembly. While one MILAbot working alone is far slower than existing techniques, with a team of 20 robots working in parallel, the system catches up to or surpasses existing automation methods at a lower cost.

“One benefit of this method is how incremental it is. You can start building, and if it turns out you need a new room, you can just add onto the structure. It is also reversible, so if your use changes, you can dissemble the voxels and change the structure,” Gershenfeld says.

The researchers also developed an interface that enables users to input or hand-design a voxelized structure. The automatic system determines the paths the MILAbots should follow for construction and sends commands to the assemblers.

The next step in this project will be a larger testbed in Bhutan, using the “super fab lab” that CBA helped set up there to replicate the robots to test construction for a planned sustainable city, Gershenfeld says.

Left to right: Yeshey Wangmo Lepcha, Tshering Wangzom, and Miana Smith stand under an arch created with voxels, as part of a working visit from the Bhutanese team to MIT. Credit: Courtesy of the researchers.

Additional areas of future work include studying the stability of voxel structures under lateral loads, improving the design tool to account for the physics of the system, enhancing the MILAbots, and evaluating voxels that have integrated sheeting, insulation, or electrical and plumbing routing.

“Our work helps support why doing this type of distributed robot assembly might be a practical way to bring digital fabrication into building construction,” Smith says.

This work was funded, in part, by the MIT Center for Bits and Atoms Consortia.

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.

In an article published today in Nature, Sony AI introduce Ace, the first robot to beat elite human players in competitive physical sport.

Although AI systems have shown advanced performance in digital domains and board games (such as complex video games, chess and Go), translating this to physical performance has remained a significant challenge. Such a feat requires perception, planning, and control to work in a high-speed domain on the scale of milliseconds. Table tennis is a demanding and complex real-world test for robotics, requiring rapid decision-making, precise physical execution, and continuous adaptation to an unpredictable opponent. The ball’s high speed, spin, and complex trajectories are central to competitive play.

Director of Sony AI in Zürich, and project lead for Ace, Peter Dürr said “this research has shown that an autonomous robot can, in fact, win at a competitive sport, matching or exceeding the reaction time and decision making of humans in a physical space. Table tennis is a game of enormous complexity that requires split-second decisions as well as speed and power. This research breakthrough highlights the potential of physical AI agents to perform real-time interactive tasks, and represents a significant step toward creating robots with broader applications in fast, precise, and real-time human interactions.”

A complete view of table tennis robot, Ace, including arm and track. Credit: Sony AI.

What new components does Ace incorporate?

Ace combines event-based vision sensors and a control system based on model-free reinforcement learning, as well as state-of-the-art high-speed robot hardware. Ace was designed with three novel components:

• A high speed perception system composed of nine active pixel sensor cameras to determine the ball’s precise 3D position, combined with three gaze control systems that use event-based vision sensor cameras, pan/tilt mirrors, and telephoto tunable lens to measure the ball’s angular velocity and spin in real time.
• A novel control system based on model-free reinforcement learning to enable rapid adaptation and decision-making without reliance on pre-programmed models.
• High-speed robotic hardware capable of executing precise, high-speed control for agile physical interaction.

Members of the Ace research team and table tennis officials pose with the robot and its human opponent, Mayuka Taira, following an official match in December 2025.

Testing Ace against elite players

For the results reported in the Nature publication, Ace was evaluated in matches against five elite players and two professional table tennis players, under International Table Tennis Federation (ITTF) regulations. Ace achieved three victories in five matches against the elite players, along with competitive performances in the remaining matches.

There were some interesting results from the evaluations, including the fact that Ace was able to return a wide range of spins, consistently achieving over 75% return rate up to spins of 450 rad/s. The control systems behind Ace also allowed for quick reaction to unusual shots, such as balls bouncing off the net. This behavior illustrates the ability of the approach to generalize to situations that are both rare and hard to model in simulation.

Following submission of the Nature manuscript, the team conducted additional competitive matches in December 2025 and March 2026, beating professional players in the process. Compared with earlier evaluations, Ace demonstrated higher shot speeds, more aggressive placement closer to the table edge, and faster-paced rallies, reflecting continued performance gains under competitive conditions.

Find out more about the project in this video from Sony AI.

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.

By Adam Zewe

MIT researchers have spent more than a decade studying techniques that enable robots to find and manipulate hidden objects by “seeing” through obstacles. Their methods utilize surface-penetrating wireless signals that reflect off concealed items.

Now, the researchers are leveraging generative artificial intelligence models to overcome a longstanding bottleneck that limited the precision of prior approaches. The result is a new method that produces more accurate shape reconstructions, which could improve a robot’s ability to reliably grasp and manipulate objects that are blocked from view.

This new technique builds a partial reconstruction of a hidden object from reflected wireless signals and fills in the missing parts of its shape using a specially trained generative AI model.

The researchers also introduced an expanded system that uses generative AI to accurately reconstruct an entire room, including all the furniture. The system utilizes wireless signals sent from one stationary radar, which reflect off humans moving in the space.  

This overcomes one key challenge of many existing methods, which require a wireless sensor to be mounted on a mobile robot to scan the environment. And unlike some popular camera-based techniques, their method preserves the privacy of people in the environment.

These innovations could enable warehouse robots to verify packed items before shipping, eliminating waste from product returns. They could also allow smart home robots to understand someone’s location in a room, improving the safety and efficiency of human-robot interaction.

“What we’ve done now is develop generative AI models that help us understand wireless reflections. This opens up a lot of interesting new applications, but technically it is also a qualitative leap in capabilities, from being able to fill in gaps we were not able to see before to being able to interpret reflections and reconstruct entire scenes,” says Fadel Adib, associate professor in the Department of Electrical Engineering and Computer Science, director of the Signal Kinetics group in the MIT Media Lab, and senior author of two papers on these techniques. “We are using AI to finally unlock wireless vision.”

Adib is joined on the first paper by lead author and research assistant Laura Dodds; as well as research assistants Maisy Lam, Waleed Akbar, and Yibo Cheng; and on the second paper by lead author and former postdoc Kaichen Zhou; Dodds; and research assistant Sayed Saad Afzal. Both papers will be presented at the IEEE Conference on Computer Vision and Pattern Recognition.

Surmounting specularity

The Adib Group previously demonstrated the use of millimeter wave (mmWave) signals to create accurate reconstructions of 3D objects that are hidden from view, like a lost wallet buried under a pile.

These waves, which are the same type of signals used in Wi-Fi, can pass through common obstructions like drywall, plastic, and cardboard, and reflect off hidden objects.

But mmWaves usually reflect in a specular manner, which means a wave reflects in a single direction after striking a surface. So large portions of the surface will reflect signals away from the mmWave sensor, making those areas effectively invisible.

“When we want to reconstruct an object, we are only able to see the top surface and we can’t see any of the bottom or sides,” Dodds explains.

The researchers previously used principles from physics to interpret reflected signals, but this limits the accuracy of the reconstructed 3D shape.

In the new papers, they overcame that limitation by using a generative AI model to fill in parts that are missing from a partial reconstruction.

“But the challenge then becomes: How do you train these models to fill in these gaps?” Adib says.

Usually, researchers use extremely large datasets to train a generative AI model, which is one reason models like Claude and Llama exhibit such impressive performance. But no mmWave datasets are large enough for training.

Instead, the researchers adapted the images in large computer vision datasets to mimic the properties in mmWave reflections.

“We were simulating the property of specularity and the noise we get from these reflections so we can apply existing datasets to our domain. It would have taken years for us to collect enough new data to do this,” Lam says.

The researchers embed the physics of mmWave reflections directly into these adapted data, creating a synthetic dataset they use to teach a generative AI model to perform plausible shape reconstructions.

The complete system, called Wave-Former, proposes a set of potential object surfaces based on mmWave reflections, feeds them to the generative AI model to complete the shape, and then refines the surfaces until it achieves a full reconstruction.

Wave-Former was able to generate faithful reconstructions of about 70 everyday objects, such as cans, boxes, utensils, and fruit, boosting accuracy by nearly 20 percent over state-of-the-art baselines. The objects were hidden behind or under cardboard, wood, drywall, plastic, and fabric.

The team also built an expanded system that fully reconstructs entire indoor scenes by leveraging wireless signal reflections off humans moving in a room. Credit: Courtesy of the researchers.

Seeing “ghosts”

The team used this same approach to build an expanded system that fully reconstructs entire indoor scenes by leveraging mmWave reflections off humans moving in a room.

Human motion generates multipath reflections. Some mmWaves reflect off the human, then reflect again off a wall or object, and then arrive back at the sensor, Dodds explains.

These secondary reflections create so-called “ghost signals,” which are reflected copies of the original signal that change location as a human moves. These ghost signals are usually discarded as noise, but they also hold information about the layout of the room.

“By analyzing how these reflections change over time, we can start to get a coarse understanding of the environment around us. But trying to directly interpret these signals is going to be limited in accuracy and resolution.” Dodds says.

They used a similar training method to teach a generative AI model to interpret those coarse scene reconstructions and understand the behavior of multipath mmWave reflections. This model fills in the gaps, refining the initial reconstruction until it completes the scene.

They tested their scene reconstruction system, called RISE, using more than 100 human trajectories captured by a single mmWave radar. On average, RISE generated reconstructions that were about twice as precise than existing techniques.

In the future, the researchers want to improve the granularity and detail in their reconstructions. They also want to build large foundation models for wireless signals, like the foundation models GPT, Claude, and Gemini for language and vision, which could open new applications.

This work is supported, in part, by the National Science Foundation (NSF), the MIT Media Lab, and Amazon.

Find out more

• Wave-Former: Through-Occlusion 3D Reconstruction via Wireless Shape Completion, Laura Dodds, Maisy Lam, Waleed Akbar, Yibo Cheng, Fadel Adib
• RISE: Single Static Radar-based Indoor Scene Understanding, Kaichen Zhou, Laura Dodds, Sayed Saad Afzal, Fadel Adib

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.

Claire chatted to Krystal Mattich from Brain Corp about trustworthy autonomous robots in public spaces.

Krystal Mattich leads global data governance, system security, and privacy compliance for Brain Corp: the world’s leading autonomy platform for commercial robotics. As Senior Director of Security, Privacy, and Risk, she is the architect of the privacy-first infrastructure that powers over 40,000 BrainOS®-enabled robots across retail, airports, education and logistics. Krystal played a central role in launching Brain Corp’s public-facing Trust Center, reinforcing the company’s commitment to data transparency, GDPR compliance, and responsible AI.

In their paper Force Aware Branch Manipulation To Assist Agricultural Tasks, which was presented at IROS 2025, Madhav Rijal, Rashik Shrestha, Trevor Smith, and Yu Gu proposed a methodology to safely manipulate branches to aid various agricultural tasks. We interviewed Madhav to find out more.

Could you give us an overview of the problem you were addressing in the paper?

Madhav Rijal (MR): Our work is motivated by StickBug [1], a multi-armed robotic system for precision pollination in greenhouse environments. One of the main challenges StickBug faces is that many flowers are partially or fully hidden within the plant canopy, making them difficult to detect and reach directly for pollination. This challenge also arises in other agricultural tasks, such as fruit harvesting, where target fruits may be occluded by surrounding branches and foliage.

To address this, we study how one robot arm can safely manipulate branches so that these occluded flowers can be brought into the field of view or reachable workspace of another robot arm. This is a challenging manipulation problem because plant branches are deformable, fragile, and vary significantly from one branch to another. In addition, unlike pick-and-place tasks, where objects move freely in space, branches remain attached to the plant, which imposes additional motion constraints during manipulation. If the robot moves a branch without accounting for these constraints and safety limits, it can apply excessive force and damage the branch.

So, the core problem we addressed in this paper is: how can a robot safely manipulate branches to reveal hidden flowers while remaining aware of interaction forces and minimizing damage?

How did your approach go about tackling the problem?

MR: Our approach [2] combines motion planning that accounts for branch constraints with real-time force feedback.

First, we generate a feasible manipulation path using an RRT* (rapidly exploring random tree) algorithm-based planner in the workspace. The planner respects the geometric constraints of the branch and the task requirements. We model branches as deformable linear objects and use a geometric heuristic to identify configurations that are safer to manipulate.

Then, during execution, we monitor the interaction force using a force sensor mounted on the manipulator. If the measured force exceeds a predefined safe threshold, the system does not continue along the same path. Instead, it re-plans the motion online and searches for an alternative path or goal configuration that can reduce branch stress while still achieving the task.

So, the key idea is that the robot does not plan only for reachability. It also adapts its motion based on the physical response of the branch during manipulation.

Madhav with the multi-armed pollination robot, StickBug.

What are the main contributions of your work?

MR: The main contributions of our work are:

1. A geometric heuristic model for branch manipulation that does not require branch-specific parameter tuning or physical probing.
2. A motion planning strategy for branch manipulation that respects both workspace and branch constraints, using the geometric heuristic to guide RRT* and incorporating online replanning based on force feedback.
3. An experimental demonstration showing that force feedback-based motion planning can protect branches from excessive force during manipulation.
4. Generalization across different branch types, since the method relies primarily on branch geometry and can adapt online to compensate for model inaccuracies.

Could you talk about the experiments that you carried out to test the approach?

MR: We evaluated the proposed method through a set of branch manipulation experiments using five different starting poses, all targeting a common goal region. Each configuration was tested 10 times, resulting in a total of 50 trials. A trial was considered successful if the robot brought the grasp point to within 5 cm of the goal point. For all trials, the planning time limit was set to 400 seconds, and the allowable interaction force range was −40 N to 40 N. Across the 50 trials, 39 were successful and 11 failed, corresponding to a success rate of about 78%. The average number of replanning attempts across all scenarios was 20.

In terms of force reduction, the results show a clear progression in safety. Constraint-aware planning reduced the manipulation force from above 100 N to below 60 N. Building on this, online force-aware replanning further reduced the force from about 60 N to below the desired 40 N threshold. This indicates that safety awareness through geometric heuristics, which model branches as deformable linear objects, together with force-aware online replanning, can effectively lower interaction forces during manipulation.

Overall, the experiments demonstrate that the proposed framework enables safer branch manipulation while maintaining task feasibility. By combining branch-constraint-aware planning with real-time force feedback, the robot can adapt its motion to reduce excessive force and minimize the risk of branch damage. These findings highlight the value of force-aware planning for practical robotic manipulation in agricultural environments.

Do you have plans to further extend this work?

MR: Yes, there are several directions for extending this work.

One current limitation is the need to define a safe force threshold in advance. In practice, different types of branches require different force limits for safe manipulation. A key direction for future work is to learn or estimate safe force thresholds automatically from branch geometry or visual cues.

Another extension is to improve grasp-point selection. Instead of only replanning after grasping, the system could also reason about the most suitable grasp point beforehand so that the required manipulation force is reduced from the start.

We are also interested in designing a compliant gripper with integrated force sensing that is better suited for manipulating delicate branches. In the longer term, we plan to integrate this method into a multi-arm agricultural robot, where one arm manipulates the branch and another performs pollination, pruning, or harvesting.

Overall, this work advances the development of agricultural robots that can actively manipulate branches to support tasks such as harvesting, pruning, and pollination. By exposing fruits, cut points, and hidden flowers within the canopy, this capability can help overcome key barriers to the broader adoption of robot-assisted agricultural technologies.

References

[1] Smith, Trevor, Madhav Rijal, Christopher Tatsch, R. Michael Butts, Jared Beard, R. Tyler Cook, Andy Chu, Jason Gross, and Yu Gu. Design of Stickbug: a six-armed precision pollination robot. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 69-75. IEEE, 2024.
[2] Rijal, Madhav, Rashik Shrestha, Trevor Smith, and Yu Gu, Force Aware Branch Manipulation To Assist Agricultural Tasks. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1217-1222. IEEE, 2025.

About Madhav

Madhav Rijal is a Ph.D. candidate in Mechanical Engineering at West Virginia University working in agricultural robotics. His research combines motion planning, optimization, multi-agent collaboration and distributed decision making to develop robotic systems for precision pollination and other plant-interaction tasks. His current work focuses on branch manipulation and safe robot operation in agricultural environments.

Claire chatted to Agata Suwala from the Manufacturing Technology Centre about leveraging robotics to make manufacturing systems more sustainable.

Agata Suwala is a Technology Manager at the Manufacturing Technology Centre, where she leads cutting-edge work in automation and robotics. With over a decade of experience in R&D, Agata specialises in developing and implementing advanced manufacturing systems—particularly for the aerospace sector—transforming complex, skill-intensive processes through automation. Her recent focus is on enabling the transition to a circular economy by leveraging automation and robotics to create sustainable, scalable technologies.

By Adam Zewe

Computer-aided design (CAD) systems are tried-and-true tools used to design many of the physical objects we use each day. But CAD software requires extensive expertise to master, and many tools incorporate such a high level of detail they don’t lend themselves to brainstorming or rapid prototyping.

In an effort to make design faster and more accessible for non-experts, researchers from MIT and elsewhere developed an AI-driven robotic assembly system that allows people to build physical objects by simply describing them in words.

Their system uses a generative AI model to build a 3D representation of an object’s geometry based on the user’s prompt. Then, a second generative AI model reasons about the desired object and figures out where different components should go, according to the object’s function and geometry.

The system can automatically build the object from a set of prefabricated parts using robotic assembly. It can also iterate on the design based on feedback from the user.

The researchers used this end-to-end system to fabricate furniture, including chairs and shelves, from two types of premade components. The components can be disassembled and reassembled at will, reducing the amount of waste generated through the fabrication process.

They evaluated these designs through a user study and found that more than 90 percent of participants preferred the objects made by their AI-driven system, as compared to different approaches.

While this work is an initial demonstration, the framework could be especially useful for rapid prototyping complex objects like aerospace components and architectural objects. In the longer term, it could be used in homes to fabricate furniture or other objects locally, without the need to have bulky products shipped from a central facility.

“Sooner or later, we want to be able to communicate and talk to a robot and AI system the same way we talk to each other to make things together. Our system is a first step toward enabling that future,” says lead author Alex Kyaw, a graduate student in the MIT departments of Electrical Engineering and Computer Science (EECS) and Architecture.

Kyaw is joined on the paper by Richa Gupta, an MIT architecture graduate student; Faez Ahmed, associate professor of mechanical engineering; Lawrence Sass, professor and chair of the Computation Group in the Department of Architecture; senior author Randall Davis, an EECS professor and member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); as well as others at Google Deepmind and Autodesk Research. The paper was recently presented at the Conference on Neural Information Processing Systems.

Generating a multicomponent design

While generative AI models are good at generating 3D representations, known as meshes, from text prompts, most do not produce uniform representations of an object’s geometry that have the component-level details needed for robotic assembly.

Separating these meshes into components is challenging for a model because assigning components depends on the geometry and functionality of the object and its parts.

The researchers tackled these challenges using a vision-language model (VLM), a powerful generative AI model that has been pre-trained to understand images and text. They task the VLM with figuring out how two types of prefabricated parts, structural components and panel components, should fit together to form an object.

“There are many ways we can put panels on a physical object, but the robot needs to see the geometry and reason over that geometry to make a decision about it. By serving as both the eyes and brain of the robot, the VLM enables the robot to do this,” Kyaw says.

A user prompts the system with text, perhaps by typing “make me a chair,” and gives it an AI-generated image of a chair to start.

Then, the VLM reasons about the chair and determines where panel components go on top of structural components, based on the functionality of many example objects it has seen before. For instance, the model can determine that the seat and backrest should have panels to have surfaces for someone sitting and leaning on the chair.

It outputs this information as text, such as “seat” or “backrest.” Each surface of the chair is then labeled with numbers, and the information is fed back to the VLM.

Then the VLM chooses the labels that correspond to the geometric parts of the chair that should receive panels on the 3D mesh to complete the design.

These six photos show the Text to robotic assembly of multi-component objects from different user prompts. Credit: Courtesy of the researchers.

Human-AI co-design

The user remains in the loop throughout this process and can refine the design by giving the model a new prompt, such as “only use panels on the backrest, not the seat.”

“The design space is very big, so we narrow it down through user feedback. We believe this is the best way to do it because people have different preferences, and building an idealized model for everyone would be impossible,” Kyaw says.

“The human‑in‑the‑loop process allows the users to steer the AI‑generated designs and have a sense of ownership in the final result,” adds Gupta.

Once the 3D mesh is finalized, a robotic assembly system builds the object using prefabricated parts. These reusable parts can be disassembled and reassembled into different configurations.

The researchers compared the results of their method with an algorithm that places panels on all horizontal surfaces that are facing up, and an algorithm that places panels randomly. In a user study, more than 90 percent of individuals preferred the designs made by their system.

They also asked the VLM to explain why it chose to put panels in those areas.

“We learned that the vision language model is able to understand some degree of the functional aspects of a chair, like leaning and sitting, to understand why it is placing panels on the seat and backrest. It isn’t just randomly spitting out these assignments,” Kyaw says.

In the future, the researchers want to enhance their system to handle more complex and nuanced user prompts, such as a table made out of glass and metal. In addition, they want to incorporate additional prefabricated components, such as gears, hinges, or other moving parts, so objects could have more functionality.

“Our hope is to drastically lower the barrier of access to design tools. We have shown that we can use generative AI and robotics to turn ideas into physical objects in a fast, accessible, and sustainable manner,” says Davis.

One of the key challenges in building robots for household or industrial settings is the need to master the control of high-degree-of-freedom systems such as mobile manipulators. Reinforcement learning has been a promising avenue for acquiring robot control policies, however, scaling to complex systems has proved tricky. In their work SLAC: Simulation-Pretrained Latent Action Space for Whole-Body Real-World RL, Jiaheng Hu, Peter Stone and Roberto Martín-Martín introduce a method that renders real-world reinforcement learning feasible for complex embodiments. We caught up with Jiaheng to find out more.

What is the topic of the research in your paper and why is it an interesting area for study?

This paper is about how robots (in particular, household robots like mobile manipulators) can autonomously acquire skills via interacting with the physical world (i.e. real-world reinforcement learning). Reinforcement learning (RL) is a general learning framework for learning from trial-and-error interaction with an environment, and has huge potential in allowing robots to learn tasks without humans hand-engineering the solution. RL for robotics is a very exciting field, as it can open possibilities for robots to self-improve in a scalable way, towards the creation of general-purpose household robots that can assist people in our everyday lives.

What were some of the issues with previous methods that your paper was trying to address?

Previously, most of the successful applications of RL to robotics were done by training entirely in simulation, then deploying the policy in the real-world directly (i.e. zero-shot sim2real). However, such a method has big limitations: on one hand, it is not very scalable, as you need to create task-specific, high-fidelity simulation environments that highly match the real-world environment that you want to deploy the robot in, and this can often take days or months for each and every task. On the other hand, some tasks are actually very hard to simulate, as they involve deformable objects and contact-rich interactions (for example, pouring water, folding clothes, wiping whiteboard). For these tasks, the simulation is often quite different from the real world. This is where real-world RL comes into play: if we can allow a robot to learn by directly interacting with the physical world, we don’t need a simulator anymore. However, while several attempts have been made towards realizing real-world RL, it is actually a very hard problem since: 1. Sample-inefficiency: RL requires a lot of samples (i.e. interaction with the environment) to learn good behavior, which is often impossible to collect in large quantities in the real-world. 2. Safety Issues: RL requires exploration, and random exploration in the real-world is often very very dangerous. The robot can break itself and will never be able to recover from that.

Could you tell us about the method (SLAC) that you’ve introduced?

So, creating high-fidelity simulations is very hard, and directly learning in the real-world is also really hard. What should we do? The key idea of SLAC is that we can use a low-fidelity simulation environment to assist subsequent real-world RL. Specifically, SLAC implements this idea in a two-step process: in the first step, SLAC learns a latent action space in simulation via unsupervised reinforcement learning. Unsupervised RL is a technique that allows the robot to explore a given environment and learn task-agnostic behaviors. In SLAC, we design a special unsupervised RL objective that encourages these behaviors to be safe and structured.

In the second step, we treat these learned behaviors as the new action space of the robot, where the robot does real-world RL for downstream tasks such as wiping whiteboards by making decisions in this new action space. Importantly, this method allow us to circumvent the two biggest problem of real-world RL: we don’t have to worry about safety issues since the new action space is pretrained to be always safe; and we can learn in a sample-efficient way because our new action space is trained to be very structured.

The robot carrying out the task of wiping a whiteboard.

How did you go about testing and evaluating your method, and what were some of the key results?

We test our methods on a real Tiago robot – a high degrees-of-freedom, bi-manual mobile manipulation, on a series of very challenging real-world tasks, including wiping a large whiteboard, cleaning a table, and sweeping trash into a bag. These tasks are challenging from three aspects: 1. They are visuo-motor tasks that require processing of high-dimensional image information. 2. They require the whole-body motion of the robot (i.e. controlling many degrees-of-freedom at the same time), and 3. They are contact-rich, which makes it hard to simulate accurately. On all of these tasks, our method allows us to learn high-performance policies (>80% success rate) within an hour of real-world interactions. By comparison, previous methods simply cannot solve the task, and often risk breaking the robot. So to summarize, previously it was simply not possible to solve these tasks via real-world RL, and our method has made it possible.

What are your plans for future work?

I think there is still a lot more to do at the intersection of RL and robotics. My eventual goal is to create truly self-improving robots that can learn entirely by themselves without any human involvement. More recently, I’ve been interested in how we can leverage foundation models such as vision-language models (VLMs) and vision-language-action models (VLAs) to further automate the self-improvement loop.

About Jiaheng

Jiaheng Hu is a 4th-year PhD student at UT-Austin, co-advised by Prof. Peter Stone and Prof. Roberto Martín-Martín. His research interest is in Robot Learning and Reinforcement Learning, with the long-term goal of developing self-improving robots that can learn and adapt autonomously in unstructured environments. Jiaheng’s work has been published at top-tier Robotics and ML venues, including CoRL, NeurIPS, RSS, and ICRA, and has earned multiple best paper nominations and awards. During his PhD, he interned at Google DeepMind and Ai2, and is a recipient of the Two Sigma PhD Fellowship.

Read the work in full

SLAC: Simulation-Pretrained Latent Action Space for Whole-Body Real-World RL, Jiaheng Hu, Peter Stone, Roberto Martín-Martín.

The Consumer Electronics Show (CES) is one of the largest trade events in the world. Every year, thousands of companies showcase their state-of-the-art technologies to over 100k attendees. It brings together global industry leaders, startups, and media, and is used to launch products and signal future tech trends.

All photos and videos in this article are credited to Henry Hickson.

CES 2026: Consumer Electronics Meets Physical AI

It doesn’t take long after landing in Las Vegas to see the signs of CES 2026. Within minutes, gleaming LED screens announce the show’s presence, and those ubiquitous two letters: AI. One flavour of AI dominated the show, from Jensen Huang’s keynote speech to even the smallest booths on the show floor: physical AI.

While I went there looking for robotics, the phrase ‘physical AI’ was reaching the ears of even the most casual visitor. Brands best known for televisions, laptops, and home appliances were talking confidently about embodied intelligence. At CES 2026, physical AI – in practice, robots – was no longer niche.

As a first-time CES attendee, the sheer scale of the event was overwhelming. Even with ruthless efficiency, it would have been impossible to speak to every robotics exhibitor at the Vegas mega-show. What follows, then, is not a comprehensive survey, but a set of themes that emerged repeatedly over the week – reinforced by either repetition or hype surrounding the exhibits. Together, they offer a snapshot of where consumer-facing robotics stands today, and where it appears to be heading next.

A quick caveat for seasoned roboticists: this consumer space is a far-cry from our daily interactions with lab-based robotics. Much of what’s on display is not state of the art by academic standards – but it is innovative in a different way, shaped by consumer expectations and manufacturability.

Outside CES at the Las Vegas Convention Center. 

Theme 1: Humanoids

It would be impossible to feature robotics at CES and not expect a wave of the biggest trend in recent robotics history: the rise of the humanoid. What once felt like a sci-fi dream is now a serious commercial and research focus, attracting intense interest from investors, companies, and the public. In some circles, the debate is whether humanoids are necessary at all, given the efficiency of other forms. At CES, however, they were possibly the most widespread and discussed representation of physical AI.

The range on display was vast. Booster showcased child-sized humanoids which deliver an education program on robotics. These were lined up in what resembled a classroom and were available to buy then and there – many of the classroom robots had bold red ‘SOLD’ signs attached. One Booster would set you back $10k.

Booster’s silicon class-bots wave in unison. 

Unitree drew crowds with a non-stop show of boxing, dancing, and acrobatics. While lacking in real-world application and still largely remote controlled, their rapid progress in motion capability is impressive. Unitree has achieved this with improved actuators and reinforcement learning. Their team stated that they added instabilities to their simulations, which makes the system more adaptable and gives confidence that each trick will run smoothly during demos. These boxing bots retail for $6k+ and are customisable for user needs.

Beijing-based Galaxea Dynamics are targeting research labs with a high-precision platform for foundation model development. Beginning with a simple robot arm with a camera and pincer gripper, they now offer a humanoid with inbuilt height adjustment for $70k. In addition, they are developing an open-source VLA as a starting point for users developing their own, task-specific setups. The result is an out-the box VLA system, reducing the barrier to entry for universities or R&D companies to develop their own twin armed VLA systems. Galaxea are proving especially popular with US institutions.

Galaxea Dynamics’ humanoid with pincer grippers.

Home robots were one of the most popular categories at CES. SwitchBot presented H1, a wheeled humanoid with arms that slide up and down. At this stage it is likely becoming clear that the definition of a humanoid is becoming somewhat blurry – aside from the friendly face and human-dimensioned arms, the H1 isn’t exactly a human form. Trained by observing humans completing household tasks, the H1 has so far mastered tidying around the house and loading the washing machine – if still at a slower pace than a human. Release dates are uncertain, but the company is aiming to undercut the competition by reaching a $10k price point. It’ll be interesting to see how this develops.

SwitchBot’s H1 Humanoid, ready to do your chores (very slowly).

The GR-3 humanoid by Shanghai-based Fourier caught my eye – its turtle-neck styled torso, emotive LED eyes and automotive-upholstered soft skin makes for a standout aesthetic. Its dextrous hands were shown off in close interactions with visitors via tic-tac-toe games, where it never lost. I was happy to pull off a tie. The reason for these close interactions is that GR-3 is targeted for healthcare providers, for social interaction and physical tasks around care homes and hospitals. Retailing at $100k, the primary customers are still research labs.

Fourier’s tic-tac-toe playing humanoid, GR3. 

An unexpected development for many was the entrance of large consumer brands into the humanoid space. LG’s CLOiD, integrated with its smart home ecosystem, embodied the idea of a “zero-labour home.” The CLOiD is a wheeled humanoid with a friendly LED face. Somewhat tentative in its movements, it was able to retrieve items from the fridge and fold a towel. This is still very much in development, but watch this space.

LG’s wheeled humanoid, CLOiD.

Paxini’s offerings sit exactly in the popular intersection between humanoid and wheeled robots, with an adjustable wheeled base, a floating torso adjustable up and down by rotating joints, and a humanoid upper body. Utilising VTLA models, they are targeted to move crates in warehouses or parts on factory floors, with several units already deployed.

GalBot presented another wheeled, VLA-focused platform, with demonstrations of crate stacking in a warehouse-style environment and customer service tasks in a mini convenience store setup. The platform features an adjustable torso with attached arms and head. The company has invested heavily in VLA development for manipulation, navigation, following, and dancing. They highlighted the challenge of tuning models for each platform, again pointing to data collection as a key bottleneck. Despite this, the team is pushing toward near-term deployment.

The aptly named company Humanoid, representing the UK contingent of roboticists, showcased a deployment-focused wheeled system in the form of the HMND 01 Alpha Wheeled platform. The large rectangular industrial base clearly signals a platform designed for real-world factory work rather than show-floor demos. This focus was reinforced by Humanoid’s recently announced partnership with Schaeffler, with whom they were exhibiting at CES. A smooth demo showed the height-variable robot moving metal rings between bins. While near-term efforts are firmly industrial, the team stated that a legged version better suited to home environments remains part of their longer-term roadmap.

Schaeffler’s second humanoid partner at CES, Agility Robotics, is also focused on production-ready systems for the factory floor – this time, with a fully legged platform. Their distinctive backward-knee design has been under development for many years, and the full-height robot on show was clearly built for industrial environments, with exposed metal structures and hard casing. Agility has no plans to enter consumer home robotics.

Agility’s humanoid. 

Perhaps most notable at CES was Boston Dynamic’s Atlas humanoid. It saw 4-5 row deep crowds for the entire show, as well as hitting global headlines. Atlas has been in development for over a decade and has become widely known through videos openly showing both successes and failures. The latest version was unveiled during the opening day. It performed an agile sequence of crouching, sweeping, and pointing motions. For the rest of the show, it performed regular pick-and-place demonstrations in a mock factory environment. The team openly discussed current limitations, noting that much of the system was teleoperated during the show, highlighting that the software stack is still under development.

Standing 1.9 metres tall with a lifting capacity of 50 kg, it is designed for heavy-duty work. Hyundai plans to focus production capacity on internal deployments through 2026, with broader manufacturing customers expected later. Long-term plans include scaling production to tens of thousands of units per year, signalling what may be one of the first large-scale deployments of humanoid robots with real industrial value.

The headline-grabbing Atlas humanoid demonstrates its impressive range of mobility.

Theme 2: Companion Robots

Being consumer focused, it is perhaps unsurprising that a standout theme was so consumer oriented, but I was surprised by the sheer number of newly released robots in this space. The strength of this theme lies in human–robot interaction, and in how consumer technology is being reshaped by robotics and AI.

One example was Loona, built by Beijing-based KEYi Tech. This pint-sized, four-wheeled robot pet was surprisingly insistent, identifying potential playmates through an onboard camera and bumping clumsily into their feet to ask for attention. Designed for all ages, the team reported strong interest from adults buying Loona as company for elderly parents. With some life-like qualities, Loona manages to evoke some of the emotional response of a household pet.

Loona’s adorable robotic pet.

KEYi also previewed a desktop robot concept that integrates with a smartphone, which provides access to cloud tools like ChatGPT and acts as a screen for expressive robotic eyes. Positioned as an assistant, the demo showed agentic capabilities, such as booking meetings and analysing screenshots. Its long-term usefulness remains unclear, but the slick interface was fun to interact with.

Next up was an unreleased robotic pet from OlloBot – two-wheeled, dog-sized robot pets with tablet-based faces that reminded me of Sid the Sloth from Ice Age. These child-friendly bots had ChatGPT integration as well as camera modes and learning programs.

Ollobots – your child’s new playmate or favourite toy?

Fourier presented a small, customisable desktop robot with wide digital eyes and head tracking that gave an eerie, Labubu-esque aesthetic.

Fourier’s cute desktop robot is watching you. 

Ludens AI offered INU, a fluffy desk robot. Marketed as a ‘desktop alien’, INU has come to Earth to remind users to take breaks and not work too hard. How could you say no to that? Alongside the cute rocking movements and occasional squeak, it was hard not to feel some attachment to this little robotic being.

The INU desktop alien robots waving hello. 

Rounding out the category was Sentigent’s DuoRover, a rugged, follow-me robot for outdoor adventures. Using a two-wheeled balancing design and vision-based following, it is intended to provide conversation, photos, and video recording on family outings.

Companion robotics is clearly a rapidly growing space. The open question is longevity: will these robots remain companions, or be relegated to people’s cupboards once the novelty wears off? Companies are betting that continual feature expansion, fueled by LLMs, will help them endure – but only time will tell.

Theme 3: Task-Driven, Non-Humanoid Robotics

While less prevalent at a consumer-focused show like CES, task-driven robots still made a strong impression. Typically found at trade shows, these more industry focused robots are used for warehouse / factory automation, or in hospitality. However, there was an interesting sprinkling at CES too and some great innovations in this space.

One of the most notable examples was from the Hyundai Motor Group, whose acquisition of Boston Dynamics made them one of the most popular exhibitors. Their MobED robot is a modular wheeled platform with four independently controlled wheels that can perform tight manoeuvres and can traverse varied terrain whilst maintaining a perfectly level loading bed. The modular approach means this can be used for a variety of uses cases, from delivery carts to ride-on vehicles. What stood out most was the build quality, which felt more akin to a luxury car than a typical robotics module. It will be interesting to see how this does against its lower cost competition – is there an appetite for luxury robotics platforms yet?

Hyundai’s 4-wheeled modular platforms greet the crowd by waving their wheels. 

At the other end of the spectrum was Nexus Patio Tech from Bolivia, showcasing delivery robots designed for simplicity and rapid route-to-market. With 20+ robots already operating on Bolivian streets, and many more in malls and restaurants, it was great to see a deployment first approach – although this may partly be thanks to a different regulatory landscape.

China’s Neolix demonstrated logistics at scale, with over 15,000 autonomous vehicles deployed across three sizes. By focusing exclusively on logistics, the team has simplified decision-making and accelerated deployment. For example, if there is a choice between crashing into a wall or a pedestrian, in logistics only you always choose the wall, as there is no passenger onboard to account for. This makes for a unique approach compared to some of the other, albeit highly sophisticated offerings on show at CES, such as passenger focused Waymo, Zoox, and NVIDIA. Pilots in Portugal are scheduled for the first half of 2026, which is a promising step – being able to comply with stringent EU regulations would unlock deployments in several other regions.

Neolix’s autonomous robot. 

Construction logistics were addressed by Gole Robotics from South Korea, who are tackling point-to-point material transport on construction sites – a use case so far relatively untouched by autonomous robotics due to its complexity. Their large material-carrying robot is designed to move pallets overnight in apartment buildings, allowing human workers to focus on skilled labour during the day. Here, the innovation lies as much in the redesigned workflow as in the robot itself.

Finally, USX Robotics showcased a modular indoor delivery platform for hotels and events, notable for having been developed from concept to production-ready system in just over a year.

This was not the only robot to be market-ready in lightning-fast time. With the availability of affordable and specific robotics focused components, open-source baseline libraries and easier access to manufacturing services from abroad, both the cost and complexity of building a new robotic system seems to be decreasing, enabling a host of new players into the industry. This raises interesting questions about how the market will evolve in the future. Differentiation may increasingly lie in identifying the right use cases and routes to market.

Summary

CES 2026 showed that humanoid and embodied AI systems still have a long way to go before delivering real-world value, particularly in homes. At the same time, there is a growing sense that the path to deployment is becoming clearer. A consensus has emerged across platforms: multi-camera perception, often wrist-mounted, paired with VLA models, is sufficient for most tasks. Increasingly, tactile hands and VTLA software are added.

There was a clear split between industrial and home-care humanoids. The most convincing offerings with the most deployments were in industry, with the firm favourite being logistics tasks in the manufacturing sector. With highly developed platforms like Atlas leading the charge, this industrial humanoid push feels closer than ever. Another general trend here was reaching the market via industry, followed by a longer-term goal of reaching people’s homes in some unspecified time frame. It was sometimes hard to tell whether that home-ready goal is genuine or to please the likes of investors and the excitable general public.

One interesting question for me is capability – at what point is a humanoid considered useful enough for the time-constrained consumer to buy it, not just for an early adopter? Currently, these platforms feel too slow to be useful enough for the average buyer. As development continues, I hope we begin to see more adaptive systems that provide genuine value.

What has changed is momentum. A spark of excitement has captured the imagination of so many at CES, now that what once seemed impossible is for sale, right in front of you. With greater access to compute, the development of foundation models like VLAS and VTLAs, a roadmap is emerging. The race now is for data, data, and more data –it’ll be a fascinating one to follow over the coming years. I left CES with the sense that the future is happening now.

* Watch a compilation of all the clips featured here and more on our youtube channel:

By Jennifer Chu

In the horticultural world, some vines are especially grabby. As they grow, the woody tendrils can wrap around obstacles with enough force to pull down entire fences and trees.

Inspired by vines’ twisty tenacity, engineers at MIT and Stanford University have developed a robotic gripper that can snake around and lift a variety of objects, including a glass vase and a watermelon, offering a gentler approach compared to conventional gripper designs. A larger version of the robo-tendrils can also safely lift a human out of bed.

The new bot consists of a pressurized box, positioned near the target object, from which long, vine-like tubes inflate and grow, like socks being turned inside out. As they extend, the vines twist and coil around the object before continuing back toward the box, where they are automatically clamped in place and mechanically wound back up to gently lift the object in a soft, sling-like grasp.

The researchers demonstrated that the vine robot can safely and stably lift a variety of heavy and fragile objects. The robot can also squeeze through tight quarters and push through clutter to reach and grasp a desired object.

The team envisions that this type of robot gripper could be used in a wide range of scenarios, from agricultural harvesting to loading and unloading heavy cargo. In the near term, the group is exploring applications in eldercare settings, where soft inflatable robotic vines could help to gently lift a person out of bed.

“Transferring a person out of bed is one of the most physically strenuous tasks that a caregiver carries out,” says Kentaro Barhydt, a PhD candidate in MIT’s Department of Mechanical Engineering. “This kind of robot can help relieve the caretaker, and can be gentler and more comfortable for the patient.”

Barhydt, along with his co-first author from Stanford, O. Godson Osele, and their colleagues, present the new robotic design in the journal Science Advances. The study’s co-authors are Harry Asada, the Ford Professor of Engineering at MIT, and Allison Okamura, the Richard W. Weiland Professor of Engineering at Stanford University, along with Sreela Kodali and Cosmia du Pasquier at Stanford University, and former MIT graduate student Chase Hartquist, now at the University of Florida, Gainesville.

Open and closed

As they extend, the vines twist and coil around the object before continuing back toward the box, where they are automatically clamped in place and mechanically wound back up to gently lift the object in a soft, sling-like grasp. Image credit: Courtesy of the researchers.

The team’s Stanford collaborators, led by Okamura, pioneered the development of soft, vine-inspired robots that grow outward from their tips. These designs are largely built from thin yet sturdy pneumatic tubes that grow and inflate with controlled air pressure. As they grow, the tubes can twist, bend, and snake their way through the environment, and squeeze through tight and cluttered spaces.

Researchers have mostly explored vine robots for use in safety inspections and search and rescue operations. But at MIT, Barhydt and Asada, whose group has developed robotic aides for the elderly, wondered whether such vine-inspired robots could address certain challenges in eldercare — specifically, the challenge of safely lifting a person out of bed. Often in nursing and rehabilitation settings, this transfer process is done with a patient lift, operated by a caretaker who must first physically move a patient onto their side, then back onto a hammock-like sheet. The caretaker straps the sheet around the patient and hooks it onto the mechanical lift, which then can gently hoist the patient out of bed, similar to suspending a hammock or sling.

The MIT and Stanford team imagined that as an alternative, a vine-like robot could gently snake under and around a patient to create its own sort of sling, without a caretaker having to physically maneuver the patient. But in order to lift the sling, the researchers realized they would have to add an element that was missing in existing vine robot designs: Essentially, they would have to close the loop.

Most vine-inspired robots are designed as “open-loop” systems, meaning they act as open-ended strings that can extend and bend in different configurations, but they are not designed to secure themselves to anything to form a closed loop. If a vine robot could be made to transform from an open loop to a closed loop, Barhydt surmised that it could make itself into a sling around the object and pull itself up, along with whatever, or whomever, it might hold.

For their new study, Barhydt, Osele, and their colleagues outline the design for a new vine-inspired robotic gripper that combines both open- and closed-loop actions. In an open-loop configuration, a robotic vine can grow and twist around an object to create a firm grasp. It can even burrow under a human lying on a bed. Once a grasp is made, the vine can continue to grow back toward and attach to its source, creating a closed loop that can then be retracted to retrieve the object.

“People might assume that in order to grab something, you just reach out and grab it,” Barhydt says. “But there are different stages, such as positioning and holding. By transforming between open and closed loops, we can achieve new levels of performance by leveraging the advantages of both forms for their respective stages.”

Gentle suspension

As a demonstration of their new open- and closed-loop concept, the team built a large-scale robotic system designed to safely lift a person up from a bed. The system comprises a set of pressurized boxes attached on either end of an overhead bar. An air pump inside the boxes slowly inflates and unfurls thin vine-like tubes that extend down toward the head and foot of a bed. The air pressure can be controlled to gently work the tubes under and around a person, before stretching back up to their respective boxes. The vines then thread through a clamping mechanism that secures the vines to each box. A winch winds the vines back up toward the boxes, gently lifting the person up in the process.

“Heavy but fragile objects, such as a human body, are difficult to grasp with the robotic hands that are available today,” Asada says. “We have developed a vine-like, growing robot gripper that can wrap around an object and suspend it gently and securely.”

“There’s an entire design space we hope this work inspires our colleagues to continue to explore,” says co-lead author Osele. “I especially look forward to the implications for patient transfer applications in health care.”

“I am very excited about future work to use robots like these for physically assisting people with mobility challenges,” adds co-author Okamura. “Soft robots can be relatively safe, low-cost, and optimally designed for specific human needs, in contrast to other approaches like humanoid robots.”

While the team’s design was motivated by challenges in eldercare, the researchers realized the new design could also be adapted to perform other grasping tasks. In addition to their large-scale system, they have built a smaller version that can attach to a commercial robotic arm. With this version, the team has shown that the vine robot can grasp and lift a variety of heavy and fragile objects, including a watermelon, a glass vase, a kettle bell, a stack of metal rods, and a playground ball. The vines can also snake through a cluttered bin to pull out a desired object.

“We think this kind of robot design can be adapted to many applications,” Barhydt says. “We are also thinking about applying this to heavy industry, and things like automating the operation of cranes at ports and warehouses.”

This work was supported, in part, by the National Science Foundation and the Ford Foundation.

If you’ve ever gone hiking, you know trails can be challenging and unpredictable. A path that was clear last week might be blocked today by a fallen tree. Poor maintenance, exposed roots, loose rocks, and uneven ground further complicate the terrain, making trails difficult for a robot to navigate autonomously. After a storm, puddles can form, mud can shift, and erosion can reshape the landscape. This was the fundamental challenge in our work: how can a robot perceive, plan, and adapt in real time to safely navigate hiking trails?

Autonomous trail navigation is not just a fun robotics problem; it has potential for real-world impact. In the United States alone, there are over 193,500 miles of trails on federal lands, with many more managed by state and local agencies. Millions of people hike these trails every year.

Robots capable of navigating trails could help with:

• Trail monitoring and maintenance
• Environmental data collection
• Search-and-rescue operations
• Assisting park staff in remote or hazardous areas

Driving off-trail introduces even more uncertainty. From an environmental perspective, leaving the trail can damage vegetation, accelerate erosion, and disturb wildlife. Still, there are moments when staying strictly on the trail is unsafe or impossible. So our question became: how can a robot get from A to B while staying on the trail when possible, and intelligently leaving it when necessary for safety?

Seeing the world two ways: geometry + semantics

Our main contribution is handling uncertainty by combining two complementary ways of understanding and mapping the environment:

• Geometric Terrain Analysis using LiDAR, which tells us about slopes, height changes, and large obstacles.
• Semantic-based terrain detection, using the robot camera images, which tells us what the robot is looking at: trail, grass, rocks, tree trunks, roots, potholes, and so on.

Geometry is great for detecting big hazards, but it struggles with small obstacles and terrain that looks geometrically similar, like sand versus firm ground, or shallow puddles versus dry soil, that are dangerous enough to get a robot stuck or damaged. Semantic perception can visually distinguish these cases, especially the trail the robot is meant to follow. However, camera-based systems are sensitive to lighting and visibility, making them unreliable on their own. By fusing geometry and semantics, we obtain a far more robust representation of what is safe to drive on.

We built a hiking trail dataset, labeling images into eight terrain classes, and trained a semantic segmentation model. Notably, the model became very good at recognizing established trails. These semantic labels were projected into 3D using depth and combined with the LiDAR based geometric terrain analysis map. Using a dual k-d tree structure, we fuse everything into a single traversability map, where each point in space has a cost representing how safe it is to traverse, prioritizing trail terrain.

The next step is deciding where the robot should go next, which we address using a hierarchical planning approach. At the global level, instead of planning a full path in a single pass, the planner operates in a receding-horizon manner, continuously replanning as the robot moves through the environment. We developed a custom RRT* that biases its search toward areas with higher traversability probability and uses the traversability values as its cost function. This makes it effective at generating intermediate waypoints. A local planner then handles motion between waypoints using precomputed arc trajectories and collision avoidance from the traversability and terrain analysis maps.

In practice, this makes the robot prefer staying on the trail, but not stubborn. If the trail ahead is blocked by a hazard, such as a large rock or a steep drop, it can temporarily route through grass or another safe area around the trail and then rejoin it once conditions improve. This behavior turns out to be crucial for real trails, where obstacles are common and rarely marked in advance.

We tested our system at the West Virginia University Core Arboretum using a Clearpath Husky robot. The video below summarizes our approach, showing the robot navigating the trail alongside the geometric traversability map, the semantic map, and the combined representation that ultimately drives planning decisions.

Overall, this work shows that robots do not need perfectly paved roads to navigate effectively. With the right combination of perception and planning, they can handle winding, messy, and unstructured hiking trails.

What is next?

There is still plenty of room for improvement. Expanding the dataset to include different seasons and trail types would increase robustness. Better handling of extreme lighting and weather conditions is another important step. On the planning side, we see opportunities to further optimize how the robot balances trail adherence against efficiency.

If you’re interested in learning more, check out our paper “Autonomous Hiking Trail Navigation via Semantic Segmentation and Geometric Analysis”. We’ve also made our dataset and code open-source. And if you’re an undergraduate student interested in contributing, keep an eye out for summer REU opportunities at West Virginia University, we’re always excited to welcome new people into robotics.

By Adam Zewe

In the future, tiny flying robots could be deployed to aid in the search for survivors trapped beneath the rubble after a devastating earthquake. Like real insects, these robots could flit through tight spaces larger robots can’t reach, while simultaneously dodging stationary obstacles and pieces of falling rubble.

So far, aerial microrobots have only been able to fly slowly along smooth trajectories, far from the swift, agile flight of real insects — until now.

MIT researchers have demonstrated aerial microrobots that can fly with speed and agility that is comparable to their biological counterparts. A collaborative team designed a new AI-based controller for the robotic bug that enabled it to follow gymnastic flight paths, such as executing continuous body flips.

With a two-part control scheme that combines high performance with computational efficiency, the robot’s speed and acceleration increased by about 450 percent and 250 percent, respectively, compared to the researchers’ best previous demonstrations.

The speedy robot was agile enough to complete 10 consecutive somersaults in 11 seconds, even when wind disturbances threatened to push it off course.

“We want to be able to use these robots in scenarios that more traditional quad copter robots would have trouble flying into, but that insects could navigate. Now, with our bioinspired control framework, the flight performance of our robot is comparable to insects in terms of speed, acceleration, and the pitching angle. This is quite an exciting step toward that future goal,” says Kevin Chen, an associate professor in the Department of Electrical Engineering and Computer Science (EECS), head of the Soft and Micro Robotics Laboratory within the Research Laboratory of Electronics (RLE), and co-senior author of a paper on the robot.

Chen is joined on the paper by co-lead authors Yi-Hsuan Hsiao, an EECS MIT graduate student; Andrea Tagliabue PhD ’24; and Owen Matteson, a graduate student in the Department of Aeronautics and Astronautics (AeroAstro); as well as EECS graduate student Suhan Kim; Tong Zhao MEng ’23; and co-senior author Jonathan P. How, the Ford Professor of Engineering in the Department of Aeronautics and Astronautics and a principal investigator in the Laboratory for Information and Decision Systems (LIDS). The research appears today in Science Advances.

An AI controller

Chen’s group has been building robotic insects for more than five years.

They recently developed a more durable version of their tiny robot, a microcassette-sized device that weighs less than a paperclip. The new version utilizes larger, flapping wings that enable more agile movements. They are powered by a set of squishy artificial muscles that flap the wings at an extremely fast rate.

But the controller — the “brain” of the robot that determines its position and tells it where to fly — was hand-tuned by a human, limiting the robot’s performance.

For the robot to fly quickly and aggressively like a real insect, it needed a more robust controller that could account for uncertainty and perform complex optimizations quickly.

Such a controller would be too computationally intensive to be deployed in real time, especially with the complicated aerodynamics of the lightweight robot.

To overcome this challenge, Chen’s group joined forces with How’s team and, together, they crafted a two-step, AI-driven control scheme that provides the robustness necessary for complex, rapid maneuvers, and the computational efficiency needed for real-time deployment.

“The hardware advances pushed the controller so there was more we could do on the software side, but at the same time, as the controller developed, there was more they could do with the hardware. As Kevin’s team demonstrates new capabilities, we demonstrate that we can utilize them,” How says.

For the first step, the team built what is known as a model-predictive controller. This type of powerful controller uses a dynamic, mathematical model to predict the behavior of the robot and plan the optimal series of actions to safely follow a trajectory.

While computationally intensive, it can plan challenging maneuvers like aerial somersaults, rapid turns, and aggressive body tilting. This high-performance planner is also designed to consider constraints on the force and torque the robot could apply, which is essential for avoiding collisions.

For instance, to perform multiple flips in a row, the robot would need to decelerate in such a way that its initial conditions are exactly right for doing the flip again.

“If small errors creep in, and you try to repeat that flip 10 times with those small errors, the robot will just crash. We need to have robust flight control,” How says.

They use this expert planner to train a “policy” based on a deep-learning model, to control the robot in real time, through a process called imitation learning. A policy is the robot’s decision-making engine, which tells the robot where and how to fly.

Essentially, the imitation-learning process compresses the powerful controller into a computationally efficient AI model that can run very fast.

The key was having a smart way to create just enough training data, which would teach the policy everything it needs to know for aggressive maneuvers.

“The robust training method is the secret sauce of this technique,” How explains.

The AI-driven policy takes robot positions as inputs and outputs control commands in real time, such as thrust force and torques.

Insect-like performance

In their experiments, this two-step approach enabled the insect-scale robot to fly 447 percent faster while exhibiting a 255 percent increase in acceleration. The robot was able to complete 10 somersaults in 11 seconds, and the tiny robot never strayed more than 4 or 5 centimeters off its planned trajectory.

“This work demonstrates that soft and microrobots, traditionally limited in speed, can now leverage advanced control algorithms to achieve agility approaching that of natural insects and larger robots, opening up new opportunities for multimodal locomotion,” says Hsiao.

The researchers were also able to demonstrate saccade movement, which occurs when insects pitch very aggressively, fly rapidly to a certain position, and then pitch the other way to stop. This rapid acceleration and deceleration help insects localize themselves and see clearly.

“This bio-mimicking flight behavior could help us in the future when we start putting cameras and sensors on board the robot,” Chen says.

Adding sensors and cameras so the microrobots can fly outdoors, without being attached to a complex motion capture system, will be a major area of future work.

The researchers also want to study how onboard sensors could help the robots avoid colliding with one another or coordinate navigation.

“For the micro-robotics community, I hope this paper signals a paradigm shift by showing that we can develop a new control architecture that is high-performing and efficient at the same time,” says Chen.

“This work is especially impressive because these robots still perform precise flips and fast turns despite the large uncertainties that come from relatively large fabrication tolerances in small-scale manufacturing, wind gusts of more than 1 meter per second, and even its power tether wrapping around the robot as it performs repeated flips,” says Sarah Bergbreiter, a professor of mechanical engineering at Carnegie Mellon University, who was not involved with this work.

“Although the controller currently runs on an external computer rather than onboard the robot, the authors demonstrate that similar, but less precise, control policies may be feasible even with the more limited computation available on an insect-scale robot. This is exciting because it points toward future insect-scale robots with agility approaching that of their biological counterparts,” she adds.

This research is funded, in part, by the National Science Foundation (NSF), the Office of Naval Research, Air Force Office of Scientific Research, MathWorks, and the Zakhartchenko Fellowship.

By Perla Maiolino, University of Oxford

Robots now see the world with an ease that once belonged only to science fiction. They can recognise objects, navigate cluttered spaces and sort thousands of parcels an hour. But ask a robot to touch something gently, safely or meaningfully, and the limits appear instantly.

As a researcher in soft robotics working on artificial skin and sensorised bodies, I’ve found that trying to give robots a sense of touch forces us to confront just how astonishingly sophisticated human touch really is.

My work began with the seemingly simple question of how robots might sense the world through their bodies. Develop tactile sensors, fully cover a machine with them, process the signals and, at first glance, you should get something like touch.

Except that human touch is nothing like a simple pressure map. Our skin contains several distinct types of mechanoreceptor, each tuned to different stimuli such as vibration, stretch or texture. Our spatial resolution is remarkably fine and, crucially, touch is active: we press, slide and adjust constantly, turning raw sensation into perception through dynamic interaction.

Engineers can sometimes mimic a fingertip-scale version of this, but reproducing it across an entire soft body, and giving a robot the ability to interpret this rich sensory flow, is a challenge of a completely different order.

Working on artificial skin also quickly reveals another insight: much of what we call “intelligence” doesn’t live solely in the brain. Biology offers striking examples – most famously, the octopus.

Octopuses distribute most of their neurons throughout their limbs. Studies of their motor behaviour show an octopus arm can generate and adapt movement patterns locally based on sensory input, with limited input from the brain.

Their soft, compliant bodies contribute directly to how they act in the world. And this kind of distributed, embodied intelligence, where behaviour emerges from the interplay of body, material and environment, is increasingly influential in robotics.

Touch also happens to be the first sense that humans develop in the womb. Developmental neuroscience shows tactile sensitivity emerging from around eight weeks of gestation, then spreading across the body during the second trimester. Long before sight or hearing function reliably, the foetus explores its surroundings through touch. This is thought to help shape how infants begin forming an understanding of weight, resistance and support – the basic physics of the world.

This distinction matters for robotics too. For decades, robots have relied heavily on cameras and lidars (a sensing method that uses pulses of light to measure distance) while avoiding physical contact. But we cannot expect machines to achieve human-level competence in the physical world if they rarely experience it through touch.

Simulation can teach a robot useful behaviour, but without real physical exploration, it risks merely deploying intelligence rather than developing it. To learn in the way humans do, robots need bodies that feel.

A ‘soft’ robot hand with tactile sensors, developed by the University of Oxford’s Soft Robotics Lab, gets to grips with an apple. Video: Oxford Robotics Institute.

One approach my group is exploring is giving robots a degree of “local intelligence” in their sensorised bodies. Humans benefit from the compliance of soft tissues: skin deforms in ways that increase grip, enhance friction and filter sensory signals before they even reach the brain. This is a form of intelligence embedded directly in the anatomy.

Research in soft robotics and morphological computation argues that the body can offload some of the brain’s workload. By building robots with soft structures and low-level processing, so they can adjust grip or posture based on tactile feedback without waiting for central commands, we hope to create machines that interact more safely and naturally with the physical world.

Healthcare is one area where this capability could make a profound difference. My group recently developed a robotic patient simulator for training occupational therapists (OTs). Students often practise on one another, which makes it difficult to learn the nuanced tactile skills involved in supporting someone safely. With real patients, trainees must balance functional and affective touch, respect personal boundaries and recognise subtle cues of pain or discomfort. Research on social and affective touch shows how important these cues are to human wellbeing.

To help trainees understand these interactions, our simulator, known as Mona, produces practical behavioural responses. For example, when an OT presses on a simulated pain point in the artificial skin, the robot reacts verbally and with a small physical “hitch” of the body to mimic discomfort.

Similarly, if the trainee tries to move a limb beyond what the simulated patient can tolerate, the robot tightens or resists, offering a realistic cue that the motion should stop. By capturing tactile interaction through artificial skin, our simulator provides feedback that has never previously been available in OT training.

Robots that care

In the future, robots with safe, sensitive bodies could help address growing pressures in social care. As populations age, many families suddenly find themselves lifting, repositioning or supporting relatives without formal training. “Care robots” would help with this, potentially meaning the family member could be cared for at home longer.

Surprisingly, progress in developing this type of robot has been much slower than early expectations suggested – even in Japan, which introduced some of the first care robot prototypes. One of the most advanced examples is Airec, a humanoid robot developed as part of the Japanese government’s Moonshot programme to assist in nursing and elderly-care tasks. This multifaceted programme, launched in 2019, seeks “ambitious R&D based on daring ideas” in order to build a “society in which human beings can be free from limitations of body, brain, space and time by 2050”.

Japan’s Airec care robot is one of the most advanced in development. Video by Global Update.

Throughout the world, though, translating research prototypes into regulated robots remains difficult. High development costs, strict safety requirements, and the absence of a clear commercial market have all slowed progress. But while the technical and regulatory barriers are substantial, they are steadily being addressed.

Robots that can safely share close physical space with people need to feel and modulate how they touch anything that comes into contact with their bodies. This whole-body sensitivity is what will distinguish the next generation of soft robots from today’s rigid machines.

We are still far from robots that can handle these intimate tasks independently. But building touch-enabled machines is already reshaping our understanding of touch. Every step toward robotic tactile intelligence highlights the extraordinary sophistication of our own bodies – and the deep connection between sensation, movement and what we call intelligence.

This article was commissioned in conjunction with the Professors’ Programme, part of Prototypes for Humanity, a global initiative that showcases and accelerates academic innovation to solve social and environmental challenges. The Conversation is the media partner of Prototypes for Humanity 2025.

Perla Maiolino, Associate Professor of Engineering Science, member of the Oxford Robotics Institute, University of Oxford

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

Generations in Dialogue: Bridging Perspectives in AI is a podcast from AAAI featuring thought-provoking discussions between AI experts, practitioners, and enthusiasts from different age groups and backgrounds. Each episode delves into how generational experiences shape views on AI, exploring the challenges, opportunities, and ethical considerations that come with the advancement of this transformative technology.

Human-robot interactions and social robotics with Professor Marynel Vázquez

In the fourth episode of this new series from AAAI, host Ella Lan chats to Professor Marynel Vázquez about what inspired her research direction, how her perspective on human-robot interactions has changed over time, robots navigating the social world, potential for using robots in education, modeling interactions as graphs, addressing misunderstandings with regards to robots in society, getting input from target users, the challenge of recognising when errors happen, making robots that adapt, and more.

About Professor Marynel Vázquez:

Marynel Vázquez is a computer scientist and roboticist whose research focuses on Human-Robot Interaction (HRI), particularly in multi-party settings. She studies social group dynamics—such as spatial behavior and social influence—in HRI, and develops perception and decision-making algorithms that enable autonomous, socially aware robot behavior. A central theme in her work is modeling interactions as graphs, allowing robots to reason about individuals, relationships, and groups simultaneously. Her interdisciplinary approach combines computer science, behavioral science, and design, and she enjoys building new robotic systems and research infrastructure to bring theoretical ideas into real-world practice.

About the host

Ella Lan, a member of the AAAI Student Committee, is the host of “Generations in Dialogue: Bridging Perspectives in AI.” She is passionate about bringing together voices across career stages to explore the evolving landscape of artificial intelligence. Ella is a student at Stanford University tentatively studying Computer Science and Psychology, and she enjoys creating spaces where technical innovation intersects with ethical reflection, human values, and societal impact. Her interests span education, healthcare, and AI ethics, with a focus on building inclusive, interdisciplinary conversations that shape the future of responsible AI.

Generations in Dialogue: Bridging Perspectives in AI is a podcast from AAAI featuring thought-provoking discussions between AI experts, practitioners, and enthusiasts from different age groups and backgrounds. Each episode delves into how generational experiences shape views on AI, exploring the challenges, opportunities, and ethical considerations that come with the advancement of this transformative technology.

Embodied AI, robotics, perception, and action with Professor Roberto Martín-Martín

In the third episode of this new series from AAAI, host Ella Lan chats to Professor Roberto Martín-Martín about taking a screwdriver to his toys as a child, how his research focus has evolved over time, how different generations interact with technology, making robots for everyone, being inspired by colleagues, advice for early-career researchers, and how machines can enhance human capabilities.

About Professor Roberto Martín-Martín:

Roberto Martín-Martín is an Assistant Professor of Computer Science at the University of Texas at Austin, where his research integrates robotics, computer vision, and machine learning to build autonomous agents capable of perceiving, learning, and acting in the real world. His work spans low-level tasks like pick-and-place and navigation to complex activities such as cooking and mobile manipulation, often drawing inspiration from human cognition and integrating insights from psychology and cognitive science. He previously worked as an AI Researcher at Salesforce AI and as a Postdoctoral Scholar at the Stanford Vision and Learning Lab with Silvio Savarese and Fei-Fei Li, leading projects in visuomotor learning, mobile manipulation, and human-robot interaction. He earned his Ph.D. and M.S. from Technische Universität Berlin under Oliver Brock and a B.S. from Universidad Politécnica de Madrid. His work has been recognized with best paper awards at RSS and ICRA, and he serves as Chair of the IEEE/RAS Technical Committee on Mobile Manipulation.

About the host

Ella Lan, a member of the AAAI Student Committee, is the host of “Generations in Dialogue: Bridging Perspectives in AI.” She is passionate about bringing together voices across career stages to explore the evolving landscape of artificial intelligence. Ella is a student at Stanford University tentatively studying Computer Science and Psychology, and she enjoys creating spaces where technical innovation intersects with ethical reflection, human values, and societal impact. Her interests span education, healthcare, and AI ethics, with a focus on building inclusive, interdisciplinary conversations that shape the future of responsible AI.

By Susan Kelley

Autonomous vehicles (AVs) have been tested as taxis for decades in San Francisco, Pittsburgh and around the world, and trucking companies have enormous incentives to adopt them.

But AV companies rarely share the crash- and safety-related data that is crucial to improving the safety of their vehicles – mostly because they have little incentive to do so.

Is AV safety data an auto company’s intellectual asset or a public good? It can be both – with a little tweaking, according to a team of Cornell researchers.

The team has created a roadmap outlining the barriers and opportunities to encourage AV companies to share the data to make AVs safer, from untangling public versus private data knowledge, to regulations to creating incentive programs.

“The core of AV market competition involves who has that crash data, because once you have that data, it’s much easier for you to train your AI to not make that error. The hope is to first make this data transparent and then use it for public good, and not just profit,” said Hauke Sandhaus, M.S. ’24, a doctoral candidate at Cornell Tech and co-author of “My Precious Crash Data,” published Oct. 16 in ACM on Human-Computer Interaction and presented at the ACM SIGCHI Conference on Computer-Supported Cooperative Work & Social Computing.

His co-authors are Qian Yang, assistant professor at the Cornell Ann S. Bowers College of Computing and Information Science; Wendy Ju, associate professor of information science and design tech at Cornell Tech, the Cornell Ann S. Bowers College of Computing and Information Science and the Jacobs Technion-Cornell Institute; and Angel Hsing-Chi Hwang, a former postdoctoral associate at Cornell and now assistant professor of communication at the University of Southern California, Annenberg.

The team interviewed 12 AV company employees who work on safety in AV design and deployment, to understand how they currently manage and share safety data, the data sharing challenges and concerns they face, and their ideal data-sharing practices.

The interviews revealed the AV companies have a surprising diversity of approaches, Sandhaus said. “Everyone really has some niche, homegrown data set, and there’s really not a lot of shared knowledge between these companies,” he said. “I expected there would be much more commonality.”

The research team discovered two key barriers to sharing data – both underscoring a lack of incentives. First, crash and safety data includes information about the machine-learning models and infrastructure that the company uses to improve safety. “Data sharing, even within a company, is political and fraught,” the team wrote in the paper. Second, the interviewees believed AV safety knowledge is private and brings their company a competitive edge. “This perspective leads them to view safety knowledge embedded in data as a contested space rather than public knowledge for social good,” the team wrote.

And U.S. and European regulations are not helping. They require only information such as the month when the crash occurred, the manufacturer and whether there were injuries. That doesn’t capture the underlying unexpected factors that often cause accidents, such as a person suddenly running onto the street, drivers violating traffic rules, extreme weather conditions or lost cargo blocking the road.

To encourage more data-sharing, it’s crucial to untangle safety knowledge from proprietary data, the researchers said. For example, AV companies could share information about the accident, but not raw video footage that would reveal the company’s technical infrastructure.

Companies could also come up with “exam questions” that AVs would have to pass in order to take the road. “If you have pedestrians coming from one side and vehicles from the other side, then you can use that as a test case that other AVs also have to pass,” Sandhaus said.

Academic institutions could act as data intermediaries with which AV companies could leverage strategic collaborations. Independent research institutions and other civic organizations have set precedents working with industry partners’ public knowledge. “There are arrangements, collaboration, patterns for higher ed to contribute to this without necessarily making the entire data set public,” Qian said.

The team also proposes standardizing AV safety assessment via more effective government regulations. For example, a federal policymaking agency could create a virtual city as a testing ground, with busy traffic intersections and pedestrian-heavy roads that every AV algorithm would have to be able to navigate, she said.

Federal regulators could encourage car companies to contribute scenarios to the testing environment. “The AV companies might say, ‘I want to put my test cases there, because my car probably has passed those tests.’ That can be a mechanism for encouraging safer vehicle development,” Yang said. “Proposing policy changes always feels a little bit distant, but I do think there are near-future policy solutions in this space.”

The research was funded by the National Science Foundation and Schmidt Sciences.

In intelligent systems, applications range from autonomous robotics to predictive maintenance problems. To control these systems, the essential aspects are captured with a model. When we design controllers for these models, we almost always face the same challenge: uncertainty. We’re rarely able to see the whole picture. Sensors are noisy, models of the system are imperfect; the world never behaves exactly as expected.

Imagine a robot navigating around an obstacle to reach a “goal” location. We abstract this scenario into a grid-like environment. A rock may block the path, but the robot doesn’t know exactly where the rock is. If it did, the problem would be reasonably easy: plan a route around it. But with uncertainty about the obstacle’s position, the robot must learn to operate safely and efficiently no matter where the rock turns out to be.

This simple story captures a much broader challenge: designing controllers that can cope with both partial observability and model uncertainty. In this blog post, I will guide you through our IJCAI 2025 paper, “Robust Finite-Memory Policy Gradients for Hidden-Model POMDPs”, where we explore designing controllers that perform reliably even when the environment may not be precisely known.

When you can’t see everything

When an agent doesn’t fully observe the state, we describe its sequential decision-making problem using a partially observable Markov decision process (POMDP). POMDPs model situations in which an agent must act, based on its policy, without full knowledge of the underlying state of the system. Instead, it receives observations that provide limited information about the underlying state. To handle that ambiguity and make better decisions, the agent needs some form of memory in its policy to remember what it has seen before. We typically represent such memory using finite-state controllers (FSCs). In contrast to neural networks, these are practical and efficient policy representations that encode internal memory states that the agent updates as it acts and observes.

From partial observability to hidden models

Many situations rarely fit a single model of the system. POMDPs capture uncertainty in observations and in the outcomes of actions, but not in the model itself. Despite their generality, POMDPs can’t capture sets of partially observable environments. In reality, there may be many plausible variations, as there are always unknowns — different obstacle positions, slightly different dynamics, or varying sensor noise. A controller for a POMDP does not generalize to perturbations of the model. In our example, the rock’s location is unknown, but we still want a controller that works across all possible locations. This is a more realistic, but also a more challenging scenario.

To capture this model uncertainty, we introduced the hidden-model POMDP (HM-POMDP). Rather than describing a single environment, an HM-POMDP represents a set of possible POMDPs that share the same structure but differ in their dynamics or rewards. An important fact is that a controller for one model is also applicable to the other models in the set.

The true environment in which the agent will ultimately operate is “hidden” in this set. This means the agent must learn a controller that performs well across all possible environments. The challenge is that the agent doesn’t just have to reason about what it can’t see but also about which environment it’s operating in.

A controller for an HM-POMDP must be robust: it should perform well across all possible environments. We measure the robustness of a controller by its robust performance: the worst-case performance over all models, providing a guaranteed lower bound on the agent’s performance in the true model. If a controller performs well even in the worst case, we can be confident it will perform acceptably on any model of the set when deployed.

Towards learning robust controllers

So, how do we design such controllers?

We developed the robust finite-memory policy gradient rfPG algorithm, an iterative approach that alternates between the following two key steps:

• Robust policy evaluation: Find the worst case. Determine the environment in the set where the current controller performs the worst.
• Policy optimization: Improve the controller for the worst case. Adjust the controller’s parameters with gradients from the current worst-case environment to improve robust performance.

Over time, the controller learns robust behavior: what to remember and how to act across the encountered environments. The iterative nature of this approach is rooted in the mathematical framework of “subgradients”. We apply these gradient-based updates, also used in reinforcement learning, to improve the controller’s robust performance. While the details are technical, the intuition is simple: iteratively optimizing the controller for the worst-case models improves its robust performance across all the environments.

Under the hood, rfPG uses formal verification techniques implemented in the tool PAYNT, exploiting structural similarities to represent large sets of models and evaluate controllers across them. Thanks to these developments, our approach scales to HM-POMDPs with many environments. In practice, this means we can reason over more than a hundred thousand models.

What is the impact?

We tested rfPG on HM-POMDPs that simulated environments with uncertainty. For example, navigation problems where obstacles or sensor errors varied between models. In these tests, rfPG produced policies that were not only more robust to these variations but also generalized better to completely unseen environments than several POMDP baselines. In practice, that implies we can render controllers robust to minor variations of the model. Recall our running example, with a robot that navigates a grid-world where the rock’s location is unknown. Excitingly, rfPG solves it near-optimally with only two memory nodes! You can see the controller below.

By integrating model-based reasoning with learning-based methods, we develop algorithms for systems that account for uncertainty rather than ignore it. While the results are promising, they come from simulated domains with discrete spaces; real-world deployment will require handling the continuous nature of various problems. Still, it’s practically relevant for high-level decision-making and trustworthy by design. In the future, we will scale up — for example, by using neural networks — and aim to handle broader classes of variations in the model, such as distributions over the unknowns.

Want to know more?

Thank you for reading! I hope you found it interesting and got a sense of our work. You can find out more about my work on marisgg.github.io and about our research group at ai-fm.org.

This blog post is based on the following IJCAI 2025 paper:

• Maris F. L. Galesloot, Roman Andriushchenko, Milan Češka, Sebastian Junges, and Nils Jansen: “Robust Finite-Memory Policy Gradients for Hidden-Model POMDPs”. In IJCAI 2025, pages 8518–8526.

For more on the techniques we used from the tool PAYNT and, more generally, about using these techniques to compute FSCs, see the paper below:

• Roman Andriushchenko, Milan Češka, Filip Macák, Sebastian Junges, Joost-Pieter Katoen: “An Oracle-Guided Approach to Constrained Policy Synthesis Under Uncertainty”. In JAIR, 2025.

If you’d like to learn more about another way of handling model uncertainty, have a look at our other papers as well. For instance, in our ECAI 2025 paper, we design robust controllers using recurrent neural networks (RNNs):

• Maris F. L. Galesloot, Marnix Suilen, Thiago D. Simão, Steven Carr, Matthijs T. J. Spaan, Ufuk Topcu, and Nils Jansen: “Pessimistic Iterative Planning with RNNs for Robust POMDPs”. In ECAI, 2025.

And in our NeurIPS 2025 paper, we study the evaluation of policies:

• Merlijn Krale, Eline M. Bovy, Maris F. L. Galesloot, Thiago D. Simão, and Nils Jansen: “On Evaluating Policies for Robust POMDPs”. In NeurIPS, 2025.

Generations in Dialogue: Bridging Perspectives in AI is a new podcast from AAAI featuring thought-provoking discussions between AI experts, practitioners, and enthusiasts from different age groups and backgrounds. Each episode delves into how generational experiences shape views on AI, exploring the challenges, opportunities, and ethical considerations that come with the advancement of this transformative technology.

Multi-agent systems and human-AI interaction with Professor Manuela Veloso

In the first episode of this new series from AAAI, host Ella Lan chats to Professor Manuela Veloso about her research journey and path into AI, the history and evolution of AI research, inter-generational collaborations, teamwork, AI to assist humans, AI in finance, advice for early-career researchers, and her excitement about connecting human knowledge with computers.

About Professor Manuela Veloso

Manuela Veloso is a renowned computer scientist and a pioneer in artificial intelligence, with influential work in multi-agent systems, robotics, and human-AI collaboration. She is currently the Head of AI Research at JPMorgan Chase, where she leads efforts to integrate AI into financial services. Previously, she was the Herbert A. Simon University Professor at Carnegie Mellon University and Head of its Machine Learning Department. Her research spans autonomous robotics, planning, machine learning, and AI systems that collaborate seamlessly with humans. A Fellow of the AAAI, IEEE, and AAAS, Professor Veloso has received numerous awards for her contributions to AI and remains a leading voice in advancing responsible and interactive AI systems.

About the host

Ella Lan, a member of the AAAI Student Committee, is the host of “Generations in Dialogue: Bridging Perspectives in AI.” She is passionate about bringing together voices across career stages to explore the evolving landscape of artificial intelligence. Ella is a student at Stanford University tentatively studying Computer Science and Psychology, and she enjoys creating spaces where technical innovation intersects with ethical reflection, human values, and societal impact. Her interests span education, healthcare, and AI ethics, with a focus on building inclusive, interdisciplinary conversations that shape the future of responsible AI.

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.

The Association for the Advancement of Artificial Intelligence (AAAI) is thrilled to announce the launch of a new podcast series, “Generations in Dialogue: Bridging Perspectives in AI.” Through this series, we aim to represent diverse perspectives across generations, fostering rich conversations about the ever changing landscape of AI. “Generations in Dialogue” aims to feature unique outlooks from AI experts, practitioners, enthusiasts of all backgrounds. Each episode will cover topics ranging from the history of AI, the latest research, to exploring the future ethical implications of its newest technologies. We will delve into how generational experiences shape views on AI, exploring the challenges, opportunities, and ethical considerations that come with the advancement of this transformative technology.

We welcome you to join us on this exciting journey. Whether you are an AI professional, researcher, or student, we hope this platform will be invaluable in discovering unique perspectives of AI with a global audience, bridging existing generational gaps in the understanding of AI and expanding your network beyond your local communities.

About the host

Ella Lan, a member of the AAAI Student Committee, is the host of “Generations in Dialogue: Bridging Perspectives in AI.” She is passionate about bringing together voices across career stages to explore the evolving landscape of artificial intelligence. Ella is a student at Stanford University tentatively studying Computer Science and Psychology, and she enjoys creating spaces where technical innovation intersects with ethical reflection, human values, and societal impact. Her interests span education, healthcare, and AI ethics, with a focus on building inclusive, interdisciplinary conversations that shape the future of responsible AI.

You can find out more about the series here, and you can also watch the first episodes on the AAAI YouTube channel.

By Adam Zewe

A robot searching for workers trapped in a partially collapsed mine shaft must rapidly generate a map of the scene and identify its location within that scene as it navigates the treacherous terrain.

Researchers have recently started building powerful machine-learning models to perform this complex task using only images from the robot’s onboard cameras, but even the best models can only process a few images at a time. In a real-world disaster where every second counts, a search-and-rescue robot would need to quickly traverse large areas and process thousands of images to complete its mission.

To overcome this problem, MIT researchers drew on ideas from both recent artificial intelligence vision models and classical computer vision to develop a new system that can process an arbitrary number of images. Their system accurately generates 3D maps of complicated scenes like a crowded office corridor in a matter of seconds. 

The AI-driven system incrementally creates and aligns smaller submaps of the scene, which it stitches together to reconstruct a full 3D map while estimating the robot’s position in real-time.

Unlike many other approaches, their technique does not require calibrated cameras or an expert to tune a complex system implementation. The simpler nature of their approach, coupled with the speed and quality of the 3D reconstructions, would make it easier to scale up for real-world applications.

Beyond helping search-and-rescue robots navigate, this method could be used to make extended reality applications for wearable devices like VR headsets or enable industrial robots to quickly find and move goods inside a warehouse.

“For robots to accomplish increasingly complex tasks, they need much more complex map representations of the world around them. But at the same time, we don’t want to make it harder to implement these maps in practice. We’ve shown that it is possible to generate an accurate 3D reconstruction in a matter of seconds with a tool that works out of the box,” says Dominic Maggio, an MIT graduate student and lead author of a paper on this method.

Maggio is joined on the paper by postdoc Hyungtae Lim and senior author Luca Carlone, associate professor in MIT’s Department of Aeronautics and Astronautics (AeroAstro), principal investigator in the Laboratory for Information and Decision Systems (LIDS), and director of the MIT SPARK Laboratory. The research will be presented at the Conference on Neural Information Processing Systems.

Mapping out a solution

For years, researchers have been grappling with an essential element of robotic navigation called simultaneous localization and mapping (SLAM). In SLAM, a robot recreates a map of its environment while orienting itself within the space.

Traditional optimization methods for this task tend to fail in challenging scenes, or they require the robot’s onboard cameras to be calibrated beforehand. To avoid these pitfalls, researchers train machine-learning models to learn this task from data.

While they are simpler to implement, even the best models can only process about 60 camera images at a time, making them infeasible for applications where a robot needs to move quickly through a varied environment while processing thousands of images.

To solve this problem, the MIT researchers designed a system that generates smaller submaps of the scene instead of the entire map. Their method “glues” these submaps together into one overall 3D reconstruction. The model is still only processing a few images at a time, but the system can recreate larger scenes much faster by stitching smaller submaps together.

“This seemed like a very simple solution, but when I first tried it, I was surprised that it didn’t work that well,” Maggio says.

Searching for an explanation, he dug into computer vision research papers from the 1980s and 1990s. Through this analysis, Maggio realized that errors in the way the machine-learning models process images made aligning submaps a more complex problem.

Traditional methods align submaps by applying rotations and translations until they line up. But these new models can introduce some ambiguity into the submaps, which makes them harder to align. For instance, a 3D submap of a one side of a room might have walls that are slightly bent or stretched. Simply rotating and translating these deformed submaps to align them doesn’t work.

“We need to make sure all the submaps are deformed in a consistent way so we can align them well with each other,” Carlone explains.

A more flexible approach

Borrowing ideas from classical computer vision, the researchers developed a more flexible, mathematical technique that can represent all the deformations in these submaps. By applying mathematical transformations to each submap, this more flexible method can align them in a way that addresses the ambiguity.

Based on input images, the system outputs a 3D reconstruction of the scene and estimates of the camera locations, which the robot would use to localize itself in the space.

“Once Dominic had the intuition to bridge these two worlds — learning-based approaches and traditional optimization methods — the implementation was fairly straightforward,” Carlone says. “Coming up with something this effective and simple has potential for a lot of applications.

Their system performed faster with less reconstruction error than other methods, without requiring special cameras or additional tools to process data. The researchers generated close-to-real-time 3D reconstructions of complex scenes like the inside of the MIT Chapel using only short videos captured on a cell phone.

The average error in these 3D reconstructions was less than 5 centimeters.

In the future, the researchers want to make their method more reliable for especially complicated scenes and work toward implementing it on real robots in challenging settings.

“Knowing about traditional geometry pays off. If you understand deeply what is going on in the model, you can get much better results and make things much more scalable,” Carlone says.

This work is supported, in part, by the U.S. National Science Foundation, U.S. Office of Naval Research, and the National Research Foundation of Korea. Carlone, currently on sabbatical as an Amazon Scholar, completed this work before he joined Amazon.

The 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025) took place from October 19 to 25, 2025 in Hangzhou, China. The programme included plenary and keynote talks, workshops, tutorials, forums, competitions, and a debate. There was also an exhibition where companies and institutions were able to showcase their latest hardware and software.

We cast an eye over the social media platforms to see what participants got up to during the week.

This week, we are participating in the IEEE/RSJ International Conference on Intelligent Robots and Systems #IROS2025 in Hangzhou #China

(IRI researchers right-left): @juliaborrassol.bsky.social, David Blanco-Mulero and Anais Garrell

#IRI #IROSHangzho

[image or embed]

— IRI-Institut de Robòtica i Informàtica Industrial (@iri-robotics.bsky.social) 24 October 2025 at 01:33

View this post on Instagram

A post shared by ARS Control (@arscontrol_lab)

Truly enjoyed discussing the consolidation of specialist and generalist approaches to physical AI at #IROS2025.

Hoping to visit Hangzhou in physical rather than digital form myself in the not too distant future – second IROS AC dinner missed in a row.

#Robotics #physicalAI

[image or embed]

— Markus Wulfmeier (@mwulfmeier.bsky.social) 20 October 2025 at 11:25

View this post on Instagram

A post shared by E80 Group (@e80group)

At #IROS2025 General Chair Professor Hesheng Wang and Program Chair Professor Yi Guo share what makes this year’s conference unique, from the inspiring location to the latest research shaping the future of intelligent robotics. youtu.be/_JzGoH7wilU

[image or embed]

— WebsEdge Science (@websedgescience.bsky.social) 23 October 2025 at 21:32

View this post on Instagram

A post shared by Tetsunari Inamura (@t_inamura)

From Hangzhou, #IROS2025 unites the brightest minds in #Robotics, #AI & intelligent systems to explore the Human–Robotics Frontier. Watch IROS TV for highlights, interviews, and a behind-the-scenes look at the labs shaping our robotic future. youtu.be/SojyPncpH1g

[image or embed]

— WebsEdge Science (@websedgescience.bsky.social) 23 October 2025 at 21:25

Impressive live demonstration by @unitreerobotics.bsky.social #G1 at the #IROS2025 conference! It really works!

[image or embed]

— Davide Scaramuzza (@davidescaramuzza.bsky.social) 23 October 2025 at 15:26

View this post on Instagram

A post shared by IEEE Robotics and Automation Society (@ieee_ras_official)

An incredibly fun demo of a modular legged wheel design that merges into a single unit by @DirectDriveTech at #IROS2025.

Currently, it is being driven by two people by hand pic.twitter.com/n8J5RCH1kG

— Chenhao Li @ IROS 2025 (@breadli428) October 22, 2025

#IROS2025 Only 16 years old, Jared K. Lepora stood out as the youngest presenter at this year’s IROS, impressively showcasing the dexterous robotic hand he designed in the Educational and Emotional Robots session this afternoon. pic.twitter.com/toOavpfk26

— IROS 2025 (@IROS2025) October 23, 2025

View this post on Instagram

A post shared by IROS 2025 (@iros2025)

View this post on Instagram

A post shared by Dennis Hong (@dennishongrobot)

View this post on Instagram

A post shared by IEEE Robotics and Automation Society (@ieee_ras_official)

View this post on Instagram

A post shared by IEEE Robotics and Automation Society (@ieee_ras_official)

By Alex Shipps

Chatbots like ChatGPT and Claude have experienced a meteoric rise in usage over the past three years because they can help you with a wide range of tasks. Whether you’re writing Shakespearean sonnets, debugging code, or need an answer to an obscure trivia question, artificial intelligence systems seem to have you covered. The source of this versatility? Billions, or even trillions, of textual data points across the internet.

Those data aren’t enough to teach a robot to be a helpful household or factory assistant, though. To understand how to handle, stack, and place various arrangements of objects across diverse environments, robots need demonstrations. You can think of robot training data as a collection of how-to videos that walk the systems through each motion of a task. Collecting these demonstrations on real robots is time-consuming and not perfectly repeatable, so engineers have created training data by generating simulations with AI (which don’t often reflect real-world physics), or tediously handcrafting each digital environment from scratch.

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Toyota Research Institute may have found a way to create the diverse, realistic training grounds robots need. Their “steerable scene generation” approach creates digital scenes of things like kitchens, living rooms, and restaurants that engineers can use to simulate lots of real-world interactions and scenarios. Trained on over 44 million 3D rooms filled with models of objects such as tables and plates, the tool places existing assets in new scenes, then refines each one into a physically accurate, lifelike environment.

Steerable scene generation creates these 3D worlds by “steering” a diffusion model — an AI system that generates a visual from random noise — toward a scene you’d find in everyday life. The researchers used this generative system to “in-paint” an environment, filling in particular elements throughout the scene. You can imagine a blank canvas suddenly turning into a kitchen scattered with 3D objects, which are gradually rearranged into a scene that imitates real-world physics. For example, the system ensures that a fork doesn’t pass through a bowl on a table — a common glitch in 3D graphics known as “clipping,” where models overlap or intersect.

How exactly steerable scene generation guides its creation toward realism, however, depends on the strategy you choose. Its main strategy is “Monte Carlo tree search” (MCTS), where the model creates a series of alternative scenes, filling them out in different ways toward a particular objective (like making a scene more physically realistic, or including as many edible items as possible). It’s used by the AI program AlphaGo to beat human opponents in Go (a game similar to chess), as the system considers potential sequences of moves before choosing the most advantageous one.

“We are the first to apply MCTS to scene generation by framing the scene generation task as a sequential decision-making process,” says MIT Department of Electrical Engineering and Computer Science (EECS) PhD student Nicholas Pfaff, who is a CSAIL researcher and a lead author on a paper presenting the work. “We keep building on top of partial scenes to produce better or more desired scenes over time. As a result, MCTS creates scenes that are more complex than what the diffusion model was trained on.”

In one particularly telling experiment, MCTS added the maximum number of objects to a simple restaurant scene. It featured as many as 34 items on a table, including massive stacks of dim sum dishes, after training on scenes with only 17 objects on average.

Steerable scene generation also allows you to generate diverse training scenarios via reinforcement learning — essentially, teaching a diffusion model to fulfill an objective by trial-and-error. After you train on the initial data, your system undergoes a second training stage, where you outline a reward (basically, a desired outcome with a score indicating how close you are to that goal). The model automatically learns to create scenes with higher scores, often producing scenarios that are quite different from those it was trained on.

Users can also prompt the system directly by typing in specific visual descriptions (like “a kitchen with four apples and a bowl on the table”). Then, steerable scene generation can bring your requests to life with precision. For example, the tool accurately followed users’ prompts at rates of 98 percent when building scenes of pantry shelves, and 86 percent for messy breakfast tables. Both marks are at least a 10 percent improvement over comparable methods like “MiDiffusion” and “DiffuScene.”

The system can also complete specific scenes via prompting or light directions (like “come up with a different scene arrangement using the same objects”). You could ask it to place apples on several plates on a kitchen table, for instance, or put board games and books on a shelf. It’s essentially “filling in the blank” by slotting items in empty spaces, but preserving the rest of a scene.

According to the researchers, the strength of their project lies in its ability to create many scenes that roboticists can actually use. “A key insight from our findings is that it’s OK for the scenes we pre-trained on to not exactly resemble the scenes that we actually want,” says Pfaff. “Using our steering methods, we can move beyond that broad distribution and sample from a ‘better’ one. In other words, generating the diverse, realistic, and task-aligned scenes that we actually want to train our robots in.”

Such vast scenes became the testing grounds where they could record a virtual robot interacting with different items. The machine carefully placed forks and knives into a cutlery holder, for instance, and rearranged bread onto plates in various 3D settings. Each simulation appeared fluid and realistic, resembling the real-world, adaptable robots steerable scene generation could help train, one day.

While the system could be an encouraging path forward in generating lots of diverse training data for robots, the researchers say their work is more of a proof of concept. In the future, they’d like to use generative AI to create entirely new objects and scenes, instead of using a fixed library of assets. They also plan to incorporate articulated objects that the robot could open or twist (like cabinets or jars filled with food) to make the scenes even more interactive.

To make their virtual environments even more realistic, Pfaff and his colleagues may incorporate real-world objects by using a library of objects and scenes pulled from images on the internet and using their previous work on “Scalable Real2Sim.” By expanding how diverse and lifelike AI-constructed robot testing grounds can be, the team hopes to build a community of users that’ll create lots of data, which could then be used as a massive dataset to teach dexterous robots different skills.

“Today, creating realistic scenes for simulation can be quite a challenging endeavor; procedural generation can readily produce a large number of scenes, but they likely won’t be representative of the environments the robot would encounter in the real world. Manually creating bespoke scenes is both time-consuming and expensive,” says Jeremy Binagia, an applied scientist at Amazon Robotics who wasn’t involved in the paper. “Steerable scene generation offers a better approach: train a generative model on a large collection of pre-existing scenes and adapt it (using a strategy such as reinforcement learning) to specific downstream applications. Compared to previous works that leverage an off-the-shelf vision-language model or focus just on arranging objects in a 2D grid, this approach guarantees physical feasibility and considers full 3D translation and rotation, enabling the generation of much more interesting scenes.”

“Steerable scene generation with post training and inference-time search provides a novel and efficient framework for automating scene generation at scale,” says Toyota Research Institute roboticist Rick Cory SM ’08, PhD ’10, who also wasn’t involved in the paper. “Moreover, it can generate ‘never-before-seen’ scenes that are deemed important for downstream tasks. In the future, combining this framework with vast internet data could unlock an important milestone towards efficient training of robots for deployment in the real world.”

Pfaff wrote the paper with senior author Russ Tedrake, the Toyota Professor of Electrical Engineering and Computer Science, Aeronautics and Astronautics, and Mechanical Engineering at MIT; a senior vice president of large behavior models at the Toyota Research Institute; and CSAIL principal investigator. Other authors were Toyota Research Institute robotics researcher Hongkai Dai SM ’12, PhD ’16; team lead and Senior Research Scientist Sergey Zakharov; and Carnegie Mellon University PhD student Shun Iwase. Their work was supported, in part, by Amazon and the Toyota Research Institute. The researchers presented their work at the Conference on Robot Learning (CoRL) in September.

By Alyssa Schaechinger

While working at NASA in 2003, Dr. Robert Ambrose, director of the Robotics and Automation Design Lab (RAD Lab), designed a robot with no fixed top or bottom. A perfect sphere, the RoboBall could not flip over, and its shape promised access to places wheeled or legged machines could not reach — from the deepest lunar crater to the uneven sands of a beach. Two of his students built the first prototype, but then Ambrose shelved the idea to focus on drivable rovers for astronauts.

When Ambrose arrived at Texas A&M University in 2021, he saw a chance to reignite his idea. With funding from the Chancellor’s Research Initiative and Governor’s University Research Initiative, Ambrose brought RoboBall back to life.

Now, two decades after the original idea, RoboBall is rolling across Texas A&M University.

Driven by graduate students Rishi Jangale and Derek Pravecek, the RAD Lab is intent on sending RoboBall, a novel spherical robot, into uncharted terrain.

Jangale and Pravecek, both Ph.D. students in the J. Mike Walker ’66 Department of Mechanical Engineering, have played a significant part in getting the ball rolling once again.

“Dr. Ambrose has given us such a cool opportunity. He gives us the chance to work on RoboBall however we want,” said Jangale, who began work on RoboBall in 2022. “We manage ourselves, and we get to take RoboBall in any direction we want.”

Pravecek echoed that sense of freedom. “We get to work as actual engineers doing engineering tasks. This research teaches us things beyond what we read in textbooks,” he said. “It really is the best of both worlds.”

Robot in an airbag

At the heart of the project is the simple concept of a “robot in an airbag.” Two versions now exist in tandem. RoboBall II, a 2-foot-diameter prototype, is tuned for trial runs, monitoring power output and control algorithms. RoboBall III has a diameter of 6 feet across and is built with plans to carry payloads such as sensors, cameras or sampling tools, for real-world missions.

Upcoming tests will continue to take RoboBall into outdoor environments. RAD Lab researchers are planning field trials on the beaches of Galveston to demonstrate a water-to-land transition, testing the robot’s buoyancy and terrain adaptability in a real-world setting.

“Traditional vehicles stall or tip over in abrupt transitions,” Jangale explained. “This robot can roll out of water onto sand without worrying about orientation. It’s going where other robots can’t.”

The factors that create the versatility of RoboBall also lead to some of its challenges. Once sealed inside its protective shell, the robot can only be accessed electronically. Any mechanical failure means disassembly and digging through layers of wiring and actuators.

“Diagnostics can be a headache,” said Pravacek. “If a motor fails or a sensor disconnects, you can’t just pop open a panel. You have to take apart the whole robot and rebuild. It’s like open-heart surgery on a rolling ball.”

RoboBall’s novelty means the team often operates without a blueprint.

“Every task is new,” Jangale said. “We’re very much on our own. There’s no literature on soft-shelled spherical robots of this size that roll themselves.”

Despite those hurdles, the students find themselves surprised every time the robot outperforms expectations.

“When it does something we didn’t think was possible, I’m always surprised,” Pravecek said. “It still feels like magic.”

Student-led innovation

The team set a new record when RoboBall II reached 20 miles per hour, roughly half its theoretical power output. “We didn’t anticipate hitting that speed so soon,” Pravecek said. “It was thrilling, and it opened up new targets. Now we’re pushing even further.”

Ambrose sees these reactions as proof that student-led innovation thrives when engineers have room to explore.

“The autonomy Rishi and Derek have is exactly what a project like this needs,” he said. “They’re not just following instructions — they’re inventing the next generation of exploration tools.”

Long-term goals include autonomous navigation and remote deployment. The team hopes to see RoboBall dispatched from a lunar lander to chart steep crater walls or launched from an unmanned drone to survey post-disaster landscapes on Earth. Each ball could map terrain, transmit data back to operators and even deploy instruments in hard-to-reach spots.

“Imagine a swarm of these balls deployed after a hurricane,” Jangale said. “They could map flooded areas, find survivors and bring back essential data — all without risking human lives.”

As the RoboBall project rolls on, student-driven research stands on full display.

“Engineering is problem solving at its purest,” Ambrose said. “Give creative minds a challenge and the freedom to explore, and you’ll see innovation roll into reality.”

Tell us a bit about your PhD – where are you studying, and what is the topic of your research?

I am pursuing my PhD at Purdue University, where my dissertation focuses on developing scalable and adaptive deep learning frameworks for human activity recognition (HAR) using wearable sensors. I was drawn to this topic because wearables have the potential to transform fields like healthcare, elderly care, and long-term activity tracking. Unlike video-based recognition, which can raise privacy concerns and requires fixed camera setups, wearables are portable, non-intrusive, and capable of continuous monitoring, making them ideal for capturing activity data in natural, real-world settings.

The central challenge my dissertation addresses is that wearable data is often noisy, inconsistent, and uncertain, depending on sensor placement, movement artifacts, and device limitations. My goal is to design deep learning models that are not only computationally efficient and interpretable but also robust to the variability of real-world data. In doing so, I aim to ensure that wearable HAR systems are both practical and trustworthy for deployment outside controlled lab environments.

This research has been supported by the Polytechnic Summer Research Grant at Purdue. Beyond my dissertation work, I contribute to the research community as a reviewer for conferences such as CoDIT, CTDIAC, and IRC, and I have been invited to review for AAAI 2026. I was also involved in community building, serving as Local Organizer and Safety Chair for the 24th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2025), and continuing as Safety Chair for AAMAS 2026.

Could you give us an overview of the research you’ve carried out so far during your PhD?

So far, my research has focused on developing a hierarchical fuzzy deep neural network that can adapt to diverse human activity recognition datasets. In my initial work, I explored a hierarchical recognition approach, where simpler activities are detected at earlier levels of the model and more complex activities are recognized at higher levels. To enhance both robustness and interpretability, I integrated fuzzy logic principles into deep learning, allowing the model to better handle uncertainty in real-world sensor data.

A key strength of this model is its simplicity and low computational cost, which makes it particularly well suited for real-time activity recognition on wearable devices. I have rigorously evaluated the framework on multiple benchmark datasets of multivariate time series and systematically compared its performance against state-of-the-art methods, where it has demonstrated both competitive accuracy and improved interpretability.

Is there an aspect of your research that has been particularly interesting?

Yes, what excites me most is discovering how different approaches can make human activity recognition both smarter and more practical. For instance, integrating fuzzy logic has been fascinating, because it allows the model to capture the natural uncertainty and variability of human movement. Instead of forcing rigid classifications, the system can reason in terms of degrees of confidence, making it more interpretable and closer to how humans actually think.

I also find the hierarchical design of my model particularly interesting. Recognizing simple activities first, and then building toward more complex behaviors, mirrors the way humans often understand actions in layers. This structure not only makes the model efficient but also provides insights into how different activities relate to one another.

Beyond methodology, what motivates me is the real-world potential. The fact that these models can run efficiently on wearables means they could eventually support personalized healthcare, elderly care, and long term activity monitoring in people’s everyday lives. And since the techniques I’m developing apply broadly to time series data, their impact could extend well beyond HAR, into areas like medical diagnostics, IoT monitoring, or even audio recognition. That sense of both depth and versatility is what makes the research especially rewarding for me.

What are your plans for building on your research so far during the PhD – what aspects will you be investigating next?

Moving forward, I plan to further enhance the scalability and adaptability of my framework so that it can effectively handle large scale datasets and support real-time applications. A major focus will be on improving both the computational efficiency and interpretability of the model, ensuring it is not only powerful but also practical for deployment in real-world scenarios.

While my current research has focused on human activity recognition, I am excited to broaden the scope to the wider domain of time series classification. I see great potential in applying my framework to areas such as sound classification, physiological signal analysis, and other time-dependent domains. This will allow me to demonstrate the generalizability and robustness of my approach across diverse applications where time-based data is critical.

In the longer term, my goal is to develop a unified, scalable model for time series analysis that balances adaptability, interpretability, and efficiency. I hope such a framework can serve as a foundation for advancing not only HAR but also a broad range of healthcare, environmental, and AI-driven applications that require real-time, data-driven decision-making.

What made you want to study AI, and in particular the area of wearables?

My interest in wearables began during my time in Paris, where I was first introduced to the potential of sensor-based monitoring in healthcare. I was immediately drawn to how discreet and non-invasive wearables are compared to video-based methods, especially for applications like elderly care and patient monitoring.

More broadly, I have always been fascinated by AI’s ability to interpret complex data and uncover meaningful patterns that can enhance human well-being. Wearables offered the perfect intersection of my interests, combining cutting-edge AI techniques with practical, real-world impact, which naturally led me to focus my research on this area.

What advice would you give to someone thinking of doing a PhD in the field?

A PhD in AI demands both technical expertise and resilience. My advice would be:

• Stay curious and adaptable, because research directions evolve quickly, and the ability to pivot or explore new ideas is invaluable.
• Investigate combining disciplines. AI benefits greatly from insights in fields like psychology, healthcare, and human-computer interaction.
• Most importantly, choose a problem you are truly passionate about. That passion will sustain you through the inevitable challenges and setbacks of the PhD journey.

Approaching your research with curiosity, openness, and genuine interest can make the PhD not just a challenge, but a deeply rewarding experience.

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

Outside of research, I’m passionate about leadership and community building. As president of the Purdue Tango Club, I grew the group from just 2 students to over 40 active members, organized weekly classes, and hosted large events with internationally recognized instructors. More importantly, I focused on creating a welcoming community where students feel connected and supported. For me, tango is more than dance, it’s a way to bring people together, bridge cultures, and balance the intensity of research with creativity and joy.

I also apply these skills in academic leadership. For example, I serve as Local Organizer and Safety Chair for the AAMAS 2025 and 2026 conferences, which has given me hands-on experience managing events, coordinating teams, and creating inclusive spaces for researchers worldwide.

About Zahra

Zahra Ghorrati is a PhD candidate and teaching assistant at Purdue University, specializing in artificial intelligence and time series classification with applications in human activity recognition. She earned her undergraduate degree in Computer Software Engineering and her master’s degree in Artificial Intelligence. Her research focuses on developing scalable and interpretable fuzzy deep learning models for wearable sensor data. She has presented her work at leading international conferences and journals, including AAMAS, PAAMS, FUZZ-IEEE, IEEE Access, System and Applied Soft Computing. She has served as a reviewer for CoDIT, CTDIAC, and IRC, and has been invited to review for AAAI 2026. Zahra also contributes to community building as Local Organizer and Safety Chair for AAMAS 2025 and 2026.

October 1 was International Women in Robotics Day, and we’re delighted to introduce this year’s edition of “Women in Robotics You Need to Know About”! Robotics is no longer confined to factories or research labs. Today, it’s helping us explore space, care for people, grow food, and connect across the globe. Behind these breakthroughs are women who lead research groups, launch startups, set safety standards, and inspire the next generation. Too often, their contributions remain under-recognized, and this list is one way we can make that work visible.

This year’s list highlights 20 women in robotics you need to know about in 2025. They are professors, engineers, founders, communicators, and project leaders. Some are early in their careers, others have already shaped the field for decades. Their work ranges from tactile sensing that gives robots a human-like sense of touch, to swarm robotics for medicine and the environment, to embodied AI that may one day live in our homes. They come from across the world, including Australia, Brazil, Canada, China, Germany, Spain, Switzerland, the United Kingdom, and the United States. Together, they show us just how wide the world of robotics really is.

We publish this list not only to celebrate their achievements, but also to counter the persistent invisibility of women in robotics. Representation matters. When women’s contributions are overlooked, we risk reinforcing the false perception that robotics is not their domain. The honorees you’ll meet here prove the opposite. They are making discoveries, leading teams, starting companies, writing the standards, and pushing the boundaries of what robots can do.

The 2025 honorees

• Heba Khamis
  Heba Khamis is lecturer at UNSW and co-founder of Contactile, start-up making tactile sensors that give robots a human sense of touch to enable them to perform difficult material handling tasks.
• Kelen Teixeira Vivaldini
  Kelen Teixeira Vivaldini is a professor at UFSCar, Brazil, researching autonomous robots, intelligent systems, and mission planning, with applications in environmental monitoring and inspection.
• Natalie Panek
  Natalie Panek is a senior engineer in systems design and works in the robotics and automation division of the space technology company, MDA. She was named on Forbes 2015 30 under 30 and one of WXN’s 2014 Top 100 award winners.
• Joelle Pineau
  Joelle Pineau is a Canadian AI researcher and robotics leader and a professor at McGill. She also served as Meta AI’s vice‑president until 2025. She co‑directs the Reasoning & Learning Lab, co‑founded SmartWheeler and Nursebot and champions reproducible research.
• Hallie Siegel
  Hallie Siegel is a science communicator and former Robohub editor, building global networks that connect researchers and the public.

• Xiaorui Zhu
  Xiaorui Zhu co-founded DJI and RoboSense and directs Galaxy AI & Robotics, with award-winning research in UAVs, autonomous driving and mobile robotics.
• Lijin Aryananda
  Lijin Aryananda is a robotics researcher with 15+ years’ experience in humanoids, automation & medical devices. At ZHAW she develops AI methods for tomography, bridging academia & industry with inclusive leadership.
• Georgia Chalvatzaki
  Georgia Chalvatzaki, professor at TU Darmstadt and head of the PEARL Lab, advances human-centric robot learning. Her work blends AI and robotics to give mobile manipulators the ability to collaborate safely and intelligently with people.
• Mar Masulli
  Mar Masulli is CEO & co-founder of BitMetrics, using AI to give robots & machines vision and reasoning. She also serves on the Spanish Robotics Association board.
• Alona Kharchenko
  Alona Kharchenko is co-founder & CTO of Devanthro, building embodied AI for homes, and was recognized by Forbes as 30 Under 30.

• Nicole Robinson
  Nicole Robinson is the co-founder of Lyro Robotics, deploying AI pick-and-pack robots for industry.
• Dimitra Gkatzia
  Dimitra Gkatzia is an associate professor at Edinburgh Napier, advancing natural language generation for human-robot interaction.
• Sabine Hauert
  Sabine Hauert is a professor at the University of Bristol, co-founder of Robohub, and a pioneer in swarm robotics for nanomedicine and the environment.
• Monica Anderson
  Monica Anderson is a professor at the University of Alabama, researching distributed autonomy and inclusive human-robot teaming.
• Shilpa Gulati
  Shilpa Gulati is an experienced engineering leader with over 15 years of experience in building and scaling teams to solve complex problems in Robotics using state-of-the-art technologies.

• Shuran Song
  Shuran Song is a Stanford professor and robotics researcher, building low-cost systems for robot perception and releasing influential open datasets.
• Kathryn Zealand
  Kathryn Zealand is co-founder of Skip, building powered clothing, “e-bikes for walking”. She spun the project out of X and has a background in theoretical physics.
• Ann Virts
  Ann Virts is a NIST project leader developing test methods for mobile and wearable robots, recognized with a U.S. DOC Bronze Medal.
• Carole Franklin
  Carole Franklin directs standards development for robotics at the Association for Advancing Automation (A3), leading ANSI & ISO robot safety work. With a background at Booz Allen & Ford, she champions safe, effective deployment of robots.
• Meghan Daley
  Meghan Daley is a NASA project manager who leads teams to develop and integrate simulations for robotic operations to prepare astronauts on the ISS and beyond.

We’ll be spotlighting five honorees each week throughout October, so stay tuned for deeper profiles and stories of their work.

Why it matters

Robotics is not just about technology; it’s about people. By showcasing these individuals, we hope to inspire the next generation, connect the community, and advance the values of diversity and inclusion in STEM

Join the conversation on social media with #WomenInRobotics and help us celebrate the people making robotics better for everyone.

The article was cross-posted from Women in Robotics. Read the original here.

Claire chatted to Frances Zhu from the Colorado School of Mines about intelligent robotic systems for space exploration.

Frances Zhu has a degree in Mechanical and Aerospace Engineering and a Ph.D. in Aerospace Engineering from Cornell University. She was previously a NASA Space Technology Research Fellow and an Assistant Research Professor in the Hawaii Institute of Geophysics and Planetology at the University of Hawaii, specialising in machine learning, dynamics, systems, and controls engineering. Since 2025, she has been an Assistant Professor at the Colorado School of Mines in the Department of Mechanical Engineering, affiliated with the Robotics program and Space Resources Program.

By Ari Koeppel, Dartmouth College

NASA’s search for evidence of past life on Mars just produced an exciting update. On Sept. 10, 2025, a team of scientists published a paper detailing the Perseverance rover’s investigation of a distinctive rock outcrop called Bright Angel on the edge of Mars’ Jezero Crater. This outcrop is notable for its light-toned rocks with striking mineral nodules and multicolored, leopard print-like splotches.

By combining data from five scientific instruments, the team determined that these nodules formed through processes that could have involved microorganisms. While this finding is not direct evidence of life, it’s a compelling discovery that planetary scientists hope to look into more closely.

Bright Angel rock surface at the Beaver Falls site on Mars shows nodules on the right and a leopard-like pattern at the center. NASA/JPL-Caltech/MSSS

To appreciate how discoveries like this one come about, it’s helpful to understand how scientists engage with rover data — that is, how planetary scientists like me use robots like Perseverance on Mars as extensions of our own senses.

Experiencing Mars through data

When you strap on a virtual reality headset, you suddenly lose your orientation to the immediate surroundings, and your awareness is transported by light and sound to a fabricated environment. For Mars scientists working on rover mission teams, something very similar occurs when rovers send back their daily downlinks of data.

Several developers, including MarsVR, Planetary Visor and Access Mars, have actually worked to build virtual Mars environments for viewing with a virtual reality headset. However, much of Mars scientists’ daily work instead involves analyzing numerical data visualized in graphs and plots. These datasets, produced by state-of-the-art sensors on Mars rovers, extend far beyond human vision and hearing.

A virtual Mars environment developed by Planetary Visor incorporates both 3D landscape data and rover instrument data as pop-up plots. Scientists typically access data without entering a virtual reality space. However, tools like this give the public a sense for how mission scientists experience their work.

Developing an intuition for interpreting these complex datasets takes years, if not entire careers. It is through this “mind-data connection” that scientists build mental models of Martian landscapes – models they then communicate to the world through scientific publications.

The robots’ tool kit: Sensors and instruments

Five primary instruments on Perseverance, aided by machine learning algorithms, helped describe the unusual rock formations at a site called Beaver Falls and the past they record.

Robotic hands: Mounted on the rover’s robotic arm are tools for blowing dust aside and abrading rock surfaces. These ensure the rover analyzes clean samples.

Cameras: Perseverance hosts 19 cameras for navigation, self-inspection and science. Five science-focused cameras played a key role in this study. These cameras captured details unseeable by human eyes, including magnified mineral textures and light in infrared wavelengths. Their images revealed that Bright Angel is a mudstone, a type of sedimentary rock formed from fine sediments deposited in water.

Spectrometers: Instruments such as SuperCam and SHERLOC – scanning habitable environments with Raman and luminescence for organics and chemicals – analyze how rocks reflect or emit light across a range of wavelengths. Think of this as taking hundreds of flash photographs of the same tiny spot, all in different “colors.” These datasets, called spectra, revealed signs of water integrated into mineral structures in the rock and traces of organic molecules: the basic building blocks of life.

Subsurface radar: RIMFAX, the radar imager for Mars subsurface experiment, uses radio waves to peer beneath Mars’ surface and map rock layers. At Beaver Falls, this showed the rocks were layered over other ancient terrains, likely due to the activity of a flowing river. Areas with persistently present water are better habitats for microbes than dry or intermittently wet locations.

X-ray chemistry: PIXL, the planetary instrument for X-ray lithochemistry, bombards rock surfaces with X-rays and observes how the rock glows or reflects them. This technique can tell researchers which elements and minerals the rock contains at a fine scale. PIXL revealed that the leopard-like spots found at Beaver Falls differed chemically from the surrounding rock. The spots resembled patterns on Earth formed by chemical reactions that are mediated by microbes underwater.

Key Perseverance Mars Rover instruments used in this analysis. NASA

Together, these instruments produce a multifaceted picture of the Martian environment. Some datasets require significant processing, and refined machine learning algorithms help the mission teams turn that information into a more intuitive description of the Jezero Crater’s setting, past and present.

The challenge of uncertainty

Despite Perseverance’s remarkable tools and processing software, uncertainty remains in the results. Science, especially when conducted remotely on another planet, is rarely black and white. In this case, the chemical signatures and mineral formations at Beaver Falls are suggestive – but not conclusive – of past life on Mars.

There actually are tools, such as mass spectrometers, that can show definitively whether a rock sample contains evidence of biological activity. However, these instruments are currently too fragile, heavy and power-intensive for Mars missions.

Fortunately, Perseverance has collected and sealed rock core samples from Beaver Falls and other promising sites in Jezero Crater with the goal of sending them back to Earth. If the current Mars sample return plan can retrieve these samples, laboratories on Earth can scrutinize them far more thoroughly than the rover was able to.

Perseverance selfie at Cheyava Falls sampling site in the Beaver Falls location. NASA/JPL-Caltech/MSSS

Investing in our robotic senses

This discovery is a testament to decades of NASA’s sustained investment in Mars exploration and the work of engineering teams that developed these instruments. Yet these investments face an uncertain future.

The White House’s budget office recently proposed cutting 47% of NASA’s science funding. Such reductions could curtail ongoing missions, including Perseverance’s continued operations, which are targeted for a 23% cut, and jeopardize future plans such as the Mars sample return campaign, among many other missions.

Perseverance represents more than a machine. It is a proxy extending humanity’s senses across millions of miles to an alien world. These robotic explorers and the NASA science programs behind them are a key part of the United States’ collective quest to answer profound questions about the universe and life beyond Earth.

Ari Koeppel, Earth Sciences Postdoctoral Scientist and Adjunct Associate, Dartmouth College

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

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.

By Jennifer Chu

Growing up in the suburban town of Spring, Texas, just outside of Houston, Erik Ballesteros couldn’t help but be drawn in by the possibilities for humans in space.

It was the early 2000s, and NASA’s space shuttle program was the main transport for astronauts to the International Space Station (ISS). Ballesteros’ hometown was less than an hour from Johnson Space Center (JSC), where NASA’s mission control center and astronaut training facility are based. And as often as they could, he and his family would drive to JSC to check out the center’s public exhibits and presentations on human space exploration.

For Ballesteros, the highlight of these visits was always the tram tour, which brings visitors to JSC’s Astronaut Training Facility. There, the public can watch astronauts test out spaceflight prototypes and practice various operations in preparation for living and working on the International Space Station.

“It was a really inspiring place to be, and sometimes we would meet astronauts when they were doing signings,” he recalls. “I’d always see the gates where the astronauts would go back into the training facility, and I would think: One day I’ll be on the other side of that gate.”

Today, Ballesteros is a PhD student in mechanical engineering at MIT, and has already made good on his childhood goal. Before coming to MIT, he interned on multiple projects at JSC, working in the training facility to help test new spacesuit materials, portable life support systems, and a propulsion system for a prototype Mars rocket. He also helped train astronauts to operate the ISS’ emergency response systems.

Those early experiences steered him to MIT, where he hopes to make a more direct impact on human spaceflight. He and his advisor, Harry Asada, are building a system that will quite literally provide helping hands to future astronauts. The system, dubbed SuperLimbs, consists of a pair of wearable robotic arms that extend out from a backpack, similar to the fictional Inspector Gadget, or Doctor Octopus (“Doc Ock,” to comic book fans). Ballesteros and Asada are designing the robotic arms to be strong enough to lift an astronaut back up if they fall. The arms could also crab-walk around a spacecraft’s exterior as an astronaut inspects or makes repairs.

Ballesteros is collaborating with engineers at the NASA Jet Propulsion Laboratory to refine the design, which he plans to introduce to astronauts at JSC in the next year or two, for practical testing and user feedback. He says his time at MIT has helped him make connections across academia and in industry that have fueled his life and work.

“Success isn’t built by the actions of one, but rather it’s built on the shoulders of many,” Ballesteros says. “Connections — ones that you not just have, but maintain — are so vital to being able to open new doors and keep great ones open.”

Getting a jumpstart

Ballesteros didn’t always seek out those connections. As a kid, he counted down the minutes until the end of school, when he could go home to play video games and watch movies, “Star Wars” being a favorite. He also loved to create and had a talent for cosplay, tailoring intricate, life-like costumes inspired by cartoon and movie characters.

In high school, he took an introductory class in engineering that challenged students to build robots from kits, that they would then pit against each other, BattleBots-style. Ballesteros built a robotic ball that moved by shifting an internal weight, similar to Star Wars’ fictional, sphere-shaped BB-8. 

“It was a good introduction, and I remember thinking, this engineering thing could be fun,” he says.

After graduating high school, Ballesteros attended the University of Texas at Austin, where he pursued a bachelor’s degree in aerospace engineering. What would typically be a four-year degree stretched into an eight-year period during which Ballesteros combined college with multiple work experiences, taking on internships at NASA and elsewhere. 

In 2013, he interned at Lockheed Martin, where he contributed to various aspects of jet engine development. That experience unlocked a number of other aerospace opportunities. After a stint at NASA’s Kennedy Space Center, he went on to Johnson Space Center, where, as part of a co-op program called Pathways, he returned every spring or summer over the next five years, to intern in various departments across the center.

While the time at JSC gave him a huge amount of practical engineering experience, Ballesteros still wasn’t sure if it was the right fit. Along with his childhood fascination with astronauts and space, he had always loved cinema and the special effects that forged them. In 2018, he took a year off from the NASA Pathways program to intern at Disney, where he spent the spring semester working as a safety engineer, performing safety checks on Disney rides and attractions.

During this time, he got to know a few people in Imagineering — the research and development group that creates, designs, and builds rides, theme parks, and attractions. That summer, the group took him on as an intern, and he worked on the animatronics for upcoming rides, which involved translating certain scenes in a Disney movie into practical, safe, and functional scenes in an attraction.

“In animation, a lot of things they do are fantastical, and it was our job to find a way to make them real,” says Ballesteros, who loved every moment of the experience and hoped to be hired as an Imagineer after the internship came to an end. But he had one year left in his undergraduate degree and had to move on.

After graduating from UT Austin in December 2019, Ballesteros accepted a position at NASA’s Jet Propulsion Laboratory in Pasadena, California. He started at JPL in February of 2020, working on some last adjustments to the Mars Perseverance rover. After a few months during which JPL shifted to remote work during the Covid pandemic, Ballesteros was assigned to a project to develop a self-diagnosing spacecraft monitoring system. While working with that team, he met an engineer who was a former lecturer at MIT. As a practical suggestion, she nudged Ballesteros to consider pursuing a master’s degree, to add more value to his CV.

“She opened up the idea of going to grad school, which I hadn’t ever considered,” he says.

Full circle

In 2021, Ballesteros arrived at MIT to begin a master’s program in mechanical engineering. In interviewing with potential advisors, he immediately hit it off with Harry Asada, the Ford Professor of Enginering and director of the d’Arbeloff Laboratory for Information Systems and Technology. Years ago, Asada had pitched JPL an idea for wearable robotic arms to aid astronauts, which they quickly turned down. But Asada held onto the idea, and proposed that Ballesteros take it on as a feasibility study for his master’s thesis.

The project would require bringing a seemingly sci-fi idea into practical, functional form, for use by astronauts in future space missions. For Ballesteros, it was the perfect challenge. SuperLimbs became the focus of his master’s degree, which he earned in 2023. His initial plan was to return to industry, degree in hand. But he chose to stay at MIT to pursue a PhD, so that he could continue his work with SuperLimbs in an environment where he felt free to explore and try new things.

“MIT is like nerd Hogwarts,” he says. “One of the dreams I had as a kid was about the first day of school, and being able to build and be creative, and it was the happiest day of my life. And at MIT, I felt like that dream became reality.”

Ballesteros and Asada are now further developing SuperLimbs. The team recently re-pitched the idea to engineers at JPL, who reconsidered, and have since struck up a partnership to help test and refine the robot. In the next year or two, Ballesteros hopes to bring a fully functional, wearable design to Johnson Space Center, where astronauts can test it out in space-simulated settings.

In addition to his formal graduate work, Ballesteros has found a way to have a bit of Imagineer-like fun. He is a member of the MIT Robotics Team, which designs, builds, and runs robots in various competitions and challenges. Within this club, Ballesteros has formed a sub-club of sorts, called the Droid Builders, that aim to build animatronic droids from popular movies and franchises.

“I thought I could use what I learned from Imagineering and teach undergrads how to build robots from the ground up,” he says. “Now we’re building a full-scale WALL-E that could be fully autonomous. It’s cool to see everything come full circle.”

RoboCup is an international scientific initiative with the goal of advancing the state of the art of intelligent robots, AI and automation. The annual RoboCup event, where teams gather from across the globe to take part in competitions across a number of leagues, this year took place in Salvador, Brazil from 15-21 July. In a series of interviews, we’ve been meeting some of the RoboCup trustees, committee members, and participants, to find out more about their respective leagues. Christoph Steup is an Executive Committee member and oversees the @Work League. Ahead of the event in Brazil, we spoke to Christoph to find out more about the @Work League, the tasks that teams need to complete, and future plans for the League.

Could you start by giving us an introduction to the @Work league?

The @Work League, along with the Logistics League, forms the Industrial League. Our goal is to mimic some of the aspects of industrial production systems. An important aspect of this is factory automization and trying to mimic the factory of the future, where you have autonomous robots building products according to customer design. In these factories of the future, a single piece would be produced individually for each customer. Factories nowadays have big conveyor belts and a lot of automization, with the tasks mostly done in the same way, and you can only build stuff efficiently if you build millions of items. We are working on building individual pieces, where automization is still possible, and even a single piece can be built effectively. But obviously, in our RoboCup competitions, we are not interested in building on a factory scale – we are doing it on a very small scale. That means our robots are typically 80 centimeters long, the largest are around 70 centimeters wide, and some of them are also 80 centimeters high. So let’s say they fit in a one metre cubed box. Also, all our operations are done on the ground. This is just for simplification because building big tables to make it more realistic would also increase the cost for RoboCup and wouldn’t give much additional value.

What our robots need to do is transport objects from different workstations. So we have a default configuration where the arena starts, and there are workstations with objects lying on them, and some of these objects need to be transported to other workstations. The robot needs to do that completely autonomously. So this is one of the special things about the @Work League, that it’s completely autonomous and there is only a single restart allowed per team. So that means the robot really needs to be reliable. One of the big differences between medium teams and very good teams is that the very good teams perform well all the time whereas the medium teams have some good runs and some bad runs.

As well as the object transportation that you mentioned, and there other tasks that the teams need to carry out?

There are some special tasks in our league, like the precision placement task where the robot needs to fit an object into a cavity that is essentially the same shape and size as the object. It’s a little bit like the game that babies do to train their dexterity.

We also have a task that is inspired by a conveyor belt, but we are using a table that is constantly turning. The robots need to grasp stuff while the table is turning. This looks a little bit silly because no one would actually put a rotating table in a factory, however this is our way of actually mimicking a conveyor belt. The conveyor belt itself would be really, really difficult to integrate into the competition, so we just abstracted that and use this rotating table to actually have the same challenge but in a more manageable way. And it’s still a very, very hard challenge.

Then there are some special challenges that we integrate. For example, that robots need to report their state back so that we can observe what the robot is doing. We also have a challenge where humans are in the loop. For example, the robot brings pieces to a certain workstation where a human is present, the human assembles the pieces, and then the human needs to give a sign to the robot, and then the robot will take the piece away and put it somewhere else. This is designed to really mimic the automated factory flow.

In the past we also had a challenge where the robots needed to open a drawer, take something out and then close the drawer again. We’ve also had tasks where the robot has to handle fragile objects, like sweets, where the robot really needed to be careful in manipulating them. So in general, what differentiates us most from the Logistics League is that we are focusing a lot on manipulation and all the difficulties that come with manipulation and unknown objects, whereas Logistics is more tailored towards large-scale logistics processes with all their optimization and planning.

I was lucky enough to attend RoboCup last year in Eindhoven and what the teams were doing was really impressive. It was also interesting to see how varied the robots were, and how teams were approaching the tasks in distinct ways, with different grabbers and so on.

Yes, this difference in approaches is related to the history of our League, which is a little bit similar to the Logistics League. The Logistics League was originally a sponsored demonstration by Festo, which is a large company from Germany that creates tools, but they also have a didactics area where they provide tools to help people understand factory optimization. The @Work League was sponsored by Kuka, the robotics company, and, in the beginning, they required every team to compete with the Kuka youBot. So this was pretty much the default platform for our league, but at some point Kuka dropped from a sponsor to just an advisor to the league, and nowadays they are not part of the league at all. So when the Kuka youBot was going out of commission, the teams searched for alternatives and now we are presented with a wide variety of robots that are competing in the league, which I personally find really cool. Now we have all these different robots, all these different approaches, and some work better in some scenarios and worse in others. So we really have a scientific approach to the problem and we are really getting some insights into how you can tackle this problem on multiple levels.

Have you noticed that some of the challenges specifically are more difficult in general for all the teams?

In 2018 we introduced a challenge of so-called arbitrary surfaces which are surfaces unknown to the teams that are put on top of the workstations. The teams need to be able to deal with these surfaces. There are two surfaces that are really, really awful for the teams – one is grass that we pretty much stole from the soccer competitions! We just thought it would be funny to try it, and it was a really interesting problem, especially for some of the grippers of the different teams. For example, the current world champion, they have a rigid gripper so they can have force feedback when they grasp. However, the grass is really difficult for them – because of their rigidity, they always grasp the grass itself and then they pull up the grass with the object. And this led to some interesting problems, like they are transporting the actual surface around and not only the object. This isn’t a problem for teams using a flexible gripper. However, on the other hand, the flexible gripper makes it really difficult to assess if you have grasped the object because you have very bad force feedback. So there are two different approaches that have their pros and cons in different scenarios.

Three robots from the @Work competition in Brazil.

Are you introducing any new tasks for this year?

Yes and no. So actually we are introducing a completely new challenge which is different from what we’ve done before. The new challenge is the so-called smart farming challenge, which is opening our league to a whole new field of applications, because we are now looking at agriculture. We are working on this with Studica, a robotics company from Canada, whose hardware we are using. We’ve already given it a try at the German Open. This challenge comes with some new specifics and one these is that the teams only get the robot shortly before the competition. So they don’t really know the robot largely in advance, and they need to assemble, program and design the robot in a very short time. To compensate for this, we reduce the amount of optimization and robustness that is necessary. Because it’s an agriculture setting, we have different objects, like fruits, that the teams need to handle. This makes it a little bit more complicated because fruits have a more arbitrary shape and different levels of ripeness that need to be detected. We also have some grapes that are hanging on a wall, which is a completely different kind of manipulation task than before, because before the teams just needed to grasp things from surfaces, but now they really need to pluck stuff from a wall in a reliable way. So this is the new challenge. It also comes with a lot of software challenges because the computational power of this robot is very limited because it only has a Raspberry Pi. For example, there is not a lot of image processing possible, especially compared to the current robots, some of which even have GPUs embedded.

Is there a particular part of the hardware or software that you’ve seen some of the biggest developments in over the last year or so?

Yeah, I think one big change I observed over the years was a switch from custom neural networks for object detection to off-the-shelf components. So pretty much all teams nowadays use YOLO networks, which you can get pre-trained and they just deploy them on GPUs that they embedded into their robots. This is also one of the reasons why the robots really grew in size over the last few years because they needed space for the larger computational power. This actually made it possible for a lot of teams to reliably detect the objects. Object detection was a big problem in the beginning of the league and nowadays it’s not really a big issue – most teams are really good at that. Sometimes they are a little bit startled with decoy objects – these are objects in the arena that are not really part of the task, and they are unknown to the teams beforehand. Sometimes they are, let’s say, evil decoys that look like an object and there’s some mismatches that the teams do, but this is becoming very rare.

I think the second big change is a switch to larger manipulators with more degrees of freedom. So in the beginning, everyone had a very small manipulator with only five degrees of freedom, which limited the operating range, and nowadays pretty much all teams have a six degree of freedom manipulator with a large range. This means that they don’t need to move their robot when they are in front of a workstation, which makes them much faster and also much more precise.

Could you talk about the future plans for the League?

There are a few things we are thinking about.

With regards to the competition itself, we had a discussion with the teams about what they are interested in doing in the future. Two things came up that they really want to have. One is mobile obstacles, so they want other objects to move autonomously through the arena. We’re in the process of creating that, in cooperation with EduArt, which is a company from Germany that also provides small educational robots. And the second thing we want to introduce is a kind of humanoid robot that the teams can use to actually handle special manipulation tasks that cannot be done simply with a robot manipulator.

In terms of creating an entry-level League, we have been working on this and one potential idea is to use the smart-farming challenge as the entry point. Through the collaboration with Studica, we can provide teams with the robot and they get to keep it after the competition. At the German Open I spoke to the Rapidly-Manufactured Rescue League about this crossover between the Rescue and the Junior Leagues, and they are very keen to collaborate.

We’re also talking with the Logistics League. Their sponsor, Festo, has dropped out of their league and now they need to reorganize. We are wondering if it would be worthwhile to bring our leagues closer together, or even fuse them together to a single RoboCup Industrial league. The Logistics League wants to do more manipulation, and the @Work League wants to do more planning, so we are closing the gap naturally between the two. However, this is just a thought at the moment – we need to see how the teams react to that.

About Christoph

Christoph Steup is an active researcher specializing in various fields of robotics, including swarm robotics, precision farming, and weather-resilient autonomous driving. He currently works at the Fraunhofer Institute for Transportation and Infrastructure Systems (IVI), where he leads the Swarm Technology Group. Prior to this role, he headed the Computational Intelligence in Robotics group at the Otto von Guericke University Magdeburg. Christoph’s involvement with RoboCup began in 2015 when he joined the robOTTO team of Otto von Guericke University as team leader. His contributions to the RoboCup community expanded as he became a member of the Technical Committee for the @Work League in 2017. In 2019, he further advanced his participation by joining the Executive Committee of the league.

In this interview series, we’re meeting some of the AAAI/SIGAI Doctoral Consortium participants to find out more about their research. In this latest interview, Haimin Hu tells us about his research on the algorithmic foundations of human-centered autonomy and his plans for future projects, and gives some advice for PhD students looking to take the next step in their career.

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

My PhD research, conducted under the supervision of Professor Jaime Fernández Fisac in the Princeton Safe Robotics Lab, focuses on the algorithmic foundations of human-centered autonomy. By integrating dynamic game theory with machine learning and safety-critical control, my work aims to ensure autonomous systems, from self-driving vehicles to drones and quadrupedal robots, are performant, verifiable, and trustworthy when deployed in human-populated space. The core principle of my PhD research is to plan robots’ motion in the joint space of both physical and information states, actively ensuring safety as they navigate uncertain, changing environments and interact with humans. Its key contribution is a unified algorithmic framework—backed by game theory—that allows robots to safely interact with their human peers, adapt to human preferences and goals, and even help humans refine their skills. Specifically, my PhD work contributes to the following areas in human-centered autonomy and multi-agent systems:

• Trustworthy human–robot interaction: Planning safe and efficient robot trajectories by closing the computation loop between physical human-robot interaction and runtime learning that reduces the robot’s uncertainty about the human.
• Verifiable neural safety analysis for complex robotic systems: Learning robust neural controllers for robots with high-dimensional dynamics; guaranteeing their training-time convergence and deployment-time safety.
• Scalable interactive planning under uncertainty: Synthesizing game-theoretic control policies for complex and uncertain human–robot systems at scale.

Was there a project (or aspect of your research) that was particularly interesting?

Safety in human-robot interaction is especially difficult to define, because it hinges on an, I’d say, almost unanswerable question: How safe is safe enough when humans might behave in arbitrary ways? To give a concrete example: Is it sufficient if an autonomous vehicle can avoid hitting a fallen cyclist 99.9% of the time? What if this rate can only be achieved by the vehicle always stopping and waiting for the human to move out of the way?

I would argue that, for trustworthy deployment of robots in human-populated space, we need to complement standard statistical methods with clear-cut robust safety assurances under a vetted set of operation conditions as well established as those of bridges, power plants, and elevators. We need runtime learning to minimize the robot’s performance loss caused by safety-enforcing maneuvers; this calls for algorithms that can reduce the robot’s inherent uncertainty induced by its human peers, for example, their intent (does a human driver want to merge, cut behind, or stay in the lane?) or response (if the robot comes closer, how will the human react?). We need to close the loop between the robot’s learning and decision-making so that it can optimize efficiency by anticipating how its ongoing interaction with the human may affect the evolving uncertainty, and ultimately, its long-term performance.

What made you want to study AI, and the area of human-centered robotic systems in particular?

I’ve been fascinated by robotics and intelligent systems since childhood, when I’d spend entire days watching sci-fi anime like Mobile Suit Gundam, Neon Genesis Evangelion, or Future GPX Cyber Formula. What captivated me wasn’t just the futuristic technology, but the vision of AI as a true partner—augmenting human abilities rather than replacing them. Cyber Formula in particular planted the idea of human-AI co-evolution in my mind: an AI co-pilot that not only helps a human driver navigate high-speed, high-stakes environments, but also adapts to the driver’s style over time, ultimately making the human a better racer and deepening mutual trust along the way. Today, during my collaboration with Toyota Research Institute (TRI), I work on human-centered robotics systems that embody this principle: designing AI systems that collaborate with people in dynamic, safety-critical settings by rapidly aligning with human intent through multimodal inputs, from physical assistance to visual cues and language feedback, bringing to life the very ideas that once lived in my childhood imagination.

You’ve landed a faculty position at Johns Hopkins University (JHU) – congratulations! Could you talk a bit about the process of job searching, and perhaps share some advice and insights for PhD students who may be at a similar stage in their career?

The job search was definitely intense but also deeply rewarding. My advice to PhD students: start thinking early about the kind of long-term impact you want to make, and act early on your application package and job talk. Also, make sure you talk to people, especially your senior colleagues and peers on the job market. I personally benefited a lot from the following resources:

• The Professor Is In
• Faculty Application Advice by Professor Sylvia Herbert
• Tips for Computer Science Faculty Applications by Professor Yisong Yue
• Research Statement by Professor Tom Silver
• Autonomy Talks (examples of great talks in broad areas of autonomy and AI)

Do you have an idea of the research projects you’ll be working on at JHU?

I wish to help create a future where humans can unquestionably embrace the presence of robots around them. Towards this vision, my lab at JHU will investigate the following topics:

• Uncertainty-aware interactive motion planning: How can robots plan safe and efficient motion by accounting for their evolving uncertainty, as well as their ability to reduce it through future interaction, sensing, communication, and learning?
• Human–AI co-evolution and co-adaptation: How can embodied AI systems learn from human teammates while helping them refine existing skills and acquire new ones in a safe, personalized manner?
• Safe human-compatible autonomy: How can autonomous systems ensure prescribed safety while remaining aligned with human values and attuned to human cognitive limitations?
• Scalable and generalizable strategic decision-making: How can multi-robot systems make safe, coordinated decisions in dynamic, human-populated environments?

How was the experience attending the AAAI Doctoral Consortium?

I had the privilege of attending the 2025 AAAI Doctoral Consortium, and it was an incredibly valuable experience. I’m especially grateful to the organizers for curating such a thoughtful and supportive environment for early-career researchers. The highlight for me was the mentoring session with Dr Ming Yin (postdoc at Princeton, now faculty at Georgia Tech CSE), whose insights on navigating the uncertain and competitive job market were both encouraging and eye-opening.

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

I am passionate about skiing. I learned to ski primarily by vision-based imitation learning from a chairlift, though I’m definitely paying the price now for poor generalization! One day, I hope to build an exoskeleton that teaches me to ski better while keeping me safe on the double black diamonds.

About Haimin

Haimin Hu is an incoming Assistant Professor of Computer Science at Johns Hopkins University, where he is also a member of the Data Science and AI Institute, the Institute for Assured Autonomy, and the Laboratory for Computational Sensing and Robotics. His research focuses on the algorithmic foundations of human-centered autonomy. He has received several awards and recognitions, including a 2025 Robotics: Science and Systems Pioneer, a 2025 Cyber-Physical Systems Rising Star, and a 2024 Human-Robot Interaction Pioneer. Additionally, he has served as an Associate Editor for IEEE Robotics and Automation Letters since his fourth year as a PhD student. He obtained a PhD in Electrical and Computer Engineering from Princeton University in 2025, an MSE in Electrical Engineering from the University of Pennsylvania in 2020, and a BE in Electronic and Information Engineering from ShanghaiTech University in 2018.

The AIhub coffee corner captures the musings of AI experts over a short conversation. This month we tackle the topic of agentic AI. Joining the conversation this time are: Sanmay Das (Virginia Tech), Tom Dietterich (Oregon State University), Sabine Hauert (University of Bristol), Sarit Kraus (Bar-Ilan University), and Michael Littman (Brown University).

Sabine Hauert: Today’s topic is agentic AI. What is it? Why is it taking off? Sanmay, perhaps you could kick off with what you noticed at AAMAS [the Autonomous Agents and Multiagent Systems conference]?

Sanmay Das: It was very interesting because obviously there’s suddenly been an enormous interest in what an agent is and in the development of agentic AI. People in the AAMAS community have been thinking about what an agent is for at least three decades. Well, longer actually, but the community itself dates back about three decades in the form of these conferences. One of the very interesting questions was about why everybody is rediscovering the wheel and rewriting these papers about what it means to be an agent, and how we should think about these agents. The way in which AI has progressed, in the sense that large language models (LLMs) are now the dominant paradigm, is almost entirely different from the way in which people have thought about agents in the AAMAS community. Obviously, there’s been a lot of machine learning and reinforcement learning work, but there’s this historical tradition of thinking about reasoning and logic where you can actually have explicit world models. Even when you’re doing game theory, or MDPs, or their variants, you have an explicit world model that allows you to specify the notion of how to encode agency. Whereas I think that’s part of the disconnect now – everything is a little bit black boxy and statistical. How do you then think about what it means to be an agent? I think in terms of the underlying notion of what it means to be an agent, there’s a lot that can be learnt from what’s been done in the agents community and in philosophy.

I also think that there are some interesting ties to thinking about emergent behaviors, and multi-agent simulation. But it’s a little bit of a Wild West out there and there are all of these papers saying we need to first define what an agent is, which is definitely rediscovering the wheel. So, at AAMAS, there was a lot of discussion of stuff like that, but also questions about what this means in this particular era, because now we suddenly have these really powerful creatures that I think nobody in the AAMAS community saw coming. Fundamentally we need to adapt what we’ve been doing in the community to take into account that these are different from how we thought intelligent agents would emerge into this more general space where they can play. We need to work out how we adapt the kinds of things that we’ve learned about negotiation, agent interaction, and agent intention, to this world. Rada Mihalcea gave a really interesting keynote talk thinking about the natural language processing (NLP) side of things and the questions there.

Sabine: Do you feel like it was a new community joining the AAMAS community, or the AAMAS community that was converting?

Sanmay: Well, there were people who were coming to AAMAS and seeing that the community has been working on this for a long time. So learning something from that was definitely the vibe that I got. But my guess is, if you go to ICML or NeurIPS, that’s very much not the vibe.

Sarit Kraus: I think they’re wasting some time. I mean, forget the “what is an agent?”, but there have been many works from the agent community for many years about coordination, collaboration, etc. I heard about one recent paper where they reinvented Contract Nets. Contract Nets were introduced in 1980, and now there is a paper about it. OK, it’s LLMs that are transferring tasks from one another and signing contracts, but if they just read the past papers, it would save their time and then they could move to more interesting research questions. Currently, they say with LLM agents that you need to divide the task into sub agents. My PhD was about building a Diplomacy player, and in my design of the player there were agents that each played a different part of a Diplomacy play – one was a strategic agent, one was a Foreign Minister, etc. And now they are talking about it again.

Michael Littman: I totally agree with Sanmay and Sarit. The way I think about it is this: this notion of “let’s build agents now that we have LLMs” to me feels a little bit like we have a new programming language like Rust++, or whatever, and we can use it to write programs that we were struggling with before. It’s true that new programming languages can make some things easier, which is great, and LLMs give us a new, powerful way to create AI systems, and that’s also great. But it’s not clear that they solve the challenges that the agents community have been grappling with for so long. So, here’s a concrete example from an article that I read yesterday. Claudius is a version of Claude and it was agentified to run a small online shop. They gave it the ability to communicate with people, post slack messages, order products, set prices on things, and people were actually doing economic exchanges with the system. At the end of the day, it was terrible. Somebody talked it into buying tungsten cubes and selling them in the store. It was just nonsense. The Anthropic people viewed the experiment as a win. They said “ohh yeah, there were definitely problems, but they’re totally fixable”. And the fixes, to me, sounded like all they’d have to do is solve the problems that the agents community has been trying to solve for the last couple of decades. That’s all, and then we’ve got it perfect. And it’s not clear to me at all that just making LLMs generically better, or smarter, or better reasoners suddenly makes all these kinds of agents questions trivial because I don’t think they are. I think they’re hard for a reason and I think you have to grapple with the hard questions to actually solve these problems. But it’s true that LLMs give us a new ability to create a system that can have a conversation. But then the system’s decision-making is just really, really bad. And so I thought that was super interesting. But we agents researchers still have jobs, that’s the good news from all this.

Sabine: My bread and butter is to design agents, in our case robots, that work together to arrive at desired emergent properties and collective behaviors. From this swarm perspective, I feel that over the past 20 years we have learned a lot of the mechanisms by which you reach consensus, the mechanisms by which you automatically design agent behaviours using machine learning to enable groups to achieve a desired collective task. We know how to make agent behaviours understandable, all that good stuff you want in an engineered system. But up until now, we’ve been profoundly lacking the individual agents’ ability to interact with the world in a way that gives you richness. So in my mind, there’s a really nice interface where the agents are more capable, so they can now do those local interactions that make them useful. But we have this whole overarching way to systematically engineer collectives that I think might make the best of both worlds. I don’t know at what point that interface happens. I guess it comes partly from every community going a little bit towards the other side. So from the swarm side, we’re trying visual language models (VLMs), we’re trying to have our robots understand using LLMs their local world to communicate with humans and with each other and get a collective awareness at a very local level of what’s happening. And then we use our swarm paradigms to be able to engineer what they do as a collective using our past research expertise. I imagine for those who are just entering this discipline they need to start from the LLMs and go up. I think it’s part of the process.

Tom Dietterich: I think a lot of it just doesn’t have anything to do with agents at all, you’re writing computer programs. People found that if you try to use a single LLM to do the whole thing, the context gets all messed up and the LLM starts having trouble interpreting it. In fact, these LLMs have a relatively small short-term memory that they can effectively use before they start getting interference among the different things in the buffer. So the engineers break the system into multiple LLM calls and chain them together, and it’s not an agent, it’s just a computer program. I don’t know how many of you have seen this system called DSPy (written by Omar Khattab)? It takes an explicit sort of software engineering perspective on things. Basically, you write a type signature for each LLM module that says “here’s what it’s going to take as input, here’s what it’s going to produce as output”, you build your system, and then DSPy automatically tunes all the prompts as a sort of compiler phase to get the system to do the right thing. I want to question whether building systems with LLMs as a software engineering exercise will branch off from the building of multi-agent systems. Because virtually all the “agentic systems” are not agents in the sense that we would call them that. They don’t have autonomy any more than a regular computer program does.

Sabine: I wonder about the anthropomorphization of this, because now that you have different agents, they’re all doing a task or a job, and all of a sudden you get articles talking about how you can replace a whole team by a set of agents. So we’re no longer replacing individual jobs, we’re now replacing teams and I wonder if this terminology also doesn’t help.

Sanmay: To be clear, this idea has existed at least since the early 90s, when there were these “soft bots” that were basically running Unix commands and they were figuring out what to do themselves. It’s really no different. What people mean when they’re talking about agents is giving a piece of code the opportunity to run its own stuff and to be able to do that in service of some kind of a goal.

I think about this in terms of economic agents, because that’s what I grew up (AKA, did my PhD) thinking about. And, do I want an agent? I could think about writing an agent that manages my (non-existent) stock portfolio. If I had enough money to have a stock portfolio, I might think about writing an agent that manages that portfolio, and that’s a reasonable notion of having autonomy, right? It has some goal, which I set, and then it goes about making decisions. If you think about the sensor-actuator framework, its actuator is that it can make trades and it can take money from my bank account in order to do so. So I think that there’s something in getting back to the basic question of “how does this agent act in the world?” and then what are the percepts that it is receiving?

I completely agree with what you were saying earlier about this question of whether the LLMs enable interactions to happen in different ways. If you look at pre-LLMs, with these agents that were doing pricing, there’s this hilarious story of how some old biology textbook ended up costing $17 million on Amazon because there were these two bots that were doing the pricing of those books at two different used book stores. One of them was a slightly higher-rated store than the other, so it would take whatever price that the lower-rated store had and push it up by 10%. Then the lower-rated store was an undercutter and it would take the current highest price and go to 99% of that price. But this just led to this spiral where suddenly that book cost $17 million. This is exactly the kind of thing that’s going to happen in this world. But the thing that I’m actually somewhat worried about, and anthropomorphising, is how these agents are going to decide on their goals.There’s an opportunity for really bad errors to come out of programming that wouldn’t be as harmful in a more constrained situation.

Tom: In the reinforcement learning literature, of course, there’s all this discussion about reward hacking and so on, but now we imagine two agents interacting with each other and hacking each other’s rewards effectively, so the whole dynamics blows up – people are just not prepared.

Sabine: The breakdown of the problem that Tom mentioned, I think there’s perhaps a real benefit to having these agents that are narrower and that as a result are perhaps more verifiable at the individual level, they maybe have clearer goals, they might be more green because we might be able to constrain what area they operate with. And then in the robotics world, we’ve been looking at collaborative awareness where narrow agents that are task-specific are aware of other agents and collectively they have some awareness of what they’re meant to be doing overall. And it’s quite anti-AGI in the sense that you have lots of narrow agents again. So part of me is wondering, are we going back to heterogeneous task-specific agents and the AGI is collective, perhaps? And so this new wave, maybe it’s anti-AGI – that would be interesting!

Tom: Well, it’s almost the only way we can hope to prove the correctness of the system, to have each component narrow enough that we can actually reason about it. That’s an interesting paradox that I was missing from Stuart Russell’s “What if we succeed?” chapter in his book, which is what if we succeed in building a broad-spectrum agent, how are we going to test it?

It does seem like it would be great to have some people from the agents community speak at the machine learning conferences and try to do some diplomatic outreach. Or maybe run some workshops at those conferences.

Sarit: I was always interested in human-agent interaction and the fact that LLMs have solved the language issue for me, I’m very excited. But the other problem that has been mentioned is still here – you need to integrate strategies and decision-making. So my model is you have LLM agents that have tools that are all sorts of algorithms that we developed and implemented and there should be several of them. But the fact that somebody solved our natural language interaction, I think this is really, really great and good for the agents community as well for the computer science community generally.

Sabine: And good for the humans. It’s a good point, the humans are agents as well in those systems.

In this interview series, we’re meeting some of the AAAI/SIGAI Doctoral Consortium participants to find out more about their research. Kate Candon is a PhD student at Yale University interested in understanding how we can create interactive agents that are more effectively able to help people. We spoke to Kate to find out more about how she is leveraging explicit and implicit feedback in human-robot interactions.

Could you start by giving us a quick introduction to the topic of your research?

I study human-robot interaction. Specifically I’m interested in how we can get robots to better learn from humans in the way that they naturally teach. Typically, a lot of work in robot learning is with a human teacher who is only tasked with giving explicit feedback to the robot, but they’re not necessarily engaged in the task. So, for example, you might have a button for “good job” and “bad job”. But we know that humans give a lot of other signals, things like facial expressions and reactions to what the robot’s doing, maybe gestures like scratching their head. It could even be something like moving an object to the side that a robot hands them – that’s implicitly saying that that was the wrong thing to hand them at that time, because they’re not using it right now. Those implicit cues are trickier, they need interpretation. However, they are a way to get additional information without adding any burden to the human user. In the past, I’ve looked at these two streams (implicit and explicit feedback) separately, but my current and future research is about combining them together. Right now, we have a framework, which we are working on improving, where we can combine the implicit and explicit feedback.

In terms of picking up on the implicit feedback, how are you doing that, what’s the mechanism? Because it sounds incredibly difficult.

It can be really hard to interpret implicit cues. People will respond differently, from person to person, culture to culture, etc. And so it’s hard to know exactly which facial reaction means good versus which facial reaction means bad.

So right now, the first version of our framework is just using human actions. Seeing what the human is doing in the task can give clues about what the robot should do. They have different action spaces, but we can find an abstraction so that we can know that if a human does an action, what the similar actions would be that the robot can do. That’s the implicit feedback right now. And then, this summer, we want to extend that to using visual cues and looking at facial reactions and gestures.

So what kind of scenarios have you been kind of testing it on?

For our current project, we use a pizza making setup. Personally I really like cooking as an example because it’s a setting where it’s easy to imagine why these things would matter. I also like that cooking has this element of recipes and there is a formula, but there’s also room for personal preferences. For example, somebody likes to put their cheese on top of the pizza, so it gets really crispy, whereas other people like to put it under the meat and veggies, so that maybe it is more melty instead of crispy. Or even, some people clean up as they go versus others who wait until the end to deal with all the dishes. Another thing that I’m really excited about is that cooking can be social. Right now, we’re just working in dyadic human-robot interactions where it’s one person and one robot, but another extension that we want to work on in the coming year is extending this to group interactions. So if we have multiple people, maybe the robot can learn not only from the person reacting to the robot, but also learn from a person reacting to another person and extrapolating what that might mean for them in the collaboration.

Could you say a bit about how the work that you did earlier in your PhD has led you to this point?

When I first started my PhD, I was really interested in implicit feedback. And I thought that I wanted to focus on learning only from implicit feedback. One of my current lab mates was focused on the EMPATHIC framework, and was looking into learning from implicit human feedback, and I really liked that work and thought it was the direction that I wanted to go into.

However, that first summer of my PhD it was during COVID and so we couldn’t really have people come into the lab to interact with robots. And so instead I did an online study where I had people play a game with a robot. We recorded their face while they were playing the game, and then we tried to see if we could predict based on just facial reactions, gaze, and head orientation if we could predict what behaviors they preferred for the agent that they were playing with in the game. We actually found that we could decently well predict which of the behaviors they preferred.

The thing that was really cool was we found how much context matters. And I think this is something that is really important for going from just a solely teacher-learner paradigm to a collaboration – context really matters. What we found is that sometimes people would have really big reactions but it wasn’t necessarily to what the agent was doing, it was to something that they had done in the game. For example, there’s this clip that I always use in talks about this. This person’s playing and she has this really noticeably confused, upset look. And so at first you might think that’s negative feedback, whatever the robot did, the robot shouldn’t have done that. But if you actually look at the context, we see that it was the first time that she lost a life in this game. For the game we made a multiplayer version of Space Invaders, and she got hit by one of the aliens and her spaceship disappeared. And so based on the context, when a human looks at that, we actually say she was just confused about what happened to her. We want to filter that out and not actually consider that when reasoning about the human’s behavior. I think that was really exciting. After that, we realized that using implicit feedback only was just so hard. That’s why I’ve taken this pivot, and now I’m more interested in combining the implicit and explicit feedback together.

You mentioned the explicit element would be more binary, like good feedback, bad feedback. Would the person-in-the-loop press a button or would the feedback be given through speech?

Right now we just have a button for good job, bad job. In an HRI paper we looked at explicit feedback only. We had the same space invaders game, but we had people come into the lab and we had a little Nao robot, a little humanoid robot, sitting on the table next to them playing the game. We made it so that the person could give positive or negative feedback during the game to the robot so that it would hopefully learn better helping behavior in the collaboration. But we found that people wouldn’t actually give that much feedback because they were focused on just trying to play the game.

And so in this work we looked at whether there are different ways we can remind the person to give feedback. You don’t want to be doing it all the time because it’ll annoy the person and maybe make them worse at the game if you’re distracting them. And also you don’t necessarily always want feedback, you just want it at useful points. The two conditions we looked at were: 1) should the robot remind someone to give feedback before or after they try a new behavior? 2) should they use an “I” versus “we” framing? For example, “remember to give feedback so I can be a better teammate” versus “remember to give feedback so we can be a better team”, things like that. And we found that the “we” framing didn’t actually make people give more feedback, but it made them feel better about the feedback they gave. They felt like it was more helpful, kind of a camaraderie building. And that was only explicit feedback, but we want to see now if we combine that with a reaction from someone, maybe that point would be a good time to ask for that explicit feedback.

You’ve already touched on this but could you tell us about the future steps you have planned for the project?

The big thing motivating a lot of my work is that I want to make it easier for robots to adapt to humans with these subjective preferences. I think in terms of objective things, like being able to pick something up and move it from here to here, we’ll get to a point where robots are pretty good. But it’s these subjective preferences that are exciting. For example, I love to cook, and so I want the robot to not do too much, just to maybe do my dishes whilst I’m cooking. But someone who hates to cook might want the robot to do all of the cooking. Those are things that, even if you have the perfect robot, it can’t necessarily know those things. And so it has to be able to adapt. And a lot of the current preference learning work is so data hungry that you have to interact with it tons and tons of times for it to be able to learn. And I just don’t think that that’s realistic for people to actually have a robot in the home. If after three days you’re still telling it “no, when you help me clean up the living room, the blankets go on the couch not the chair” or something, you’re going to stop using the robot. I’m hoping that this combination of explicit and implicit feedback will help it be more naturalistic. You don’t have to necessarily know exactly the right way to give explicit feedback to get the robot to do what you want it to do. Hopefully through all of these different signals, the robot will be able to hone in a little bit faster.

I think a big future step (that is not necessarily in the near future) is incorporating language. It’s very exciting with how large language models have gotten so much better, but also there’s a lot of interesting questions. Up until now, I haven’t really included natural language. Part of it is because I’m not fully sure where it fits in the implicit versus explicit delineation. On the one hand, you can say “good job robot”, but the way you say it can mean different things – the tone is very important. For example, if you say it with a sarcastic tone, it doesn’t necessarily mean that the robot actually did a good job. So, language doesn’t fit neatly into one of the buckets, and I’m interested in future work to think more about that. I think it’s a super rich space, and it’s a way for humans to be much more granular and specific in their feedback in a natural way.

What was it that inspired you to go into this area then?

Honestly, it was a little accidental. I studied math and computer science in undergrad. After that, I worked in consulting for a couple of years and then in the public healthcare sector, for the Massachusetts Medicaid office. I decided I wanted to go back to academia and to get into AI. At the time, I wanted to combine AI with healthcare, so I was initially thinking about clinical machine learning. I’m at Yale, and there was only one person at the time doing that, so I was looking at the rest of the department and then I found Scaz (Brian Scassellati) who does a lot of work with robots for people with autism and is now moving more into robots for people with behavioral health challenges, things like dementia or anxiety. I thought his work was super interesting. I didn’t even realize that that kind of work was an option. He was working with Marynel Vázquez, a professor at Yale who was also doing human-robot interaction. She didn’t have any healthcare projects, but I interviewed with her and the questions that she was thinking about were exactly what I wanted to work on. I also really wanted to work with her. So, I accidentally stumbled into it, but I feel very grateful because I think it’s a way better fit for me than the clinical machine learning would have necessarily been. It combines a lot of what I’m interested in, and I also feel it allows me to flex back and forth between the mathy, more technical work, but then there’s also the human element, which is also super interesting and exciting to me.

Have you got any advice you’d give to someone thinking of doing a PhD in the field? Your perspective will be particularly interesting because you’ve worked outside of academia and then come back to start your PhD.

One thing is that, I mean it’s kind of cliche, but it’s not too late to start. I was hesitant because I’d been out of the field for a while, but I think if you can find the right mentor, it can be a really good experience. I think the biggest thing is finding a good advisor who you think is working on interesting questions, but also someone that you want to learn from. I feel very lucky with Marynel, she’s been a fabulous advisor. I’ve worked pretty closely with Scaz as well and they both foster this excitement about the work, but also care about me as a person. I’m not just a cog in the research machine.

The other thing I’d say is to find a lab where you have flexibility if your interests change, because it is a long time to be working on a set of projects.

For our final question, have you got an interesting non-AI related fact about you?

My main summertime hobby is playing golf. My whole family is into it – for my grandma’s 100th birthday party we had a family golf outing where we had about 40 of us golfing. And actually, that summer, when my grandma was 99, she had a par on one of the par threes – she’s my golfing role model!

About Kate

Kate Candon is a PhD candidate at Yale University in the Computer Science Department, advised by Professor Marynel Vázquez. She studies human-robot interaction, and is particularly interested in enabling robots to better learn from natural human feedback so that they can become better collaborators. She was selected for the AAMAS Doctoral Consortium in 2023 and HRI Pioneers in 2024. Before starting in human-robot interaction, she received her B.S. in Mathematics with Computer Science from MIT and then worked in consulting and in government healthcare.

RoboCup2025 took place from 15-21 July in Salvador, Brazil. The event saw around 3000 participants competing in the various leagues. In our first social media round-up post we saw what the teams got up to during the first couple of days of the event. In this second post, we take a look at the action from the final days when the competitions reached their climax.

View this post on Instagram

A post shared by RoboCup 2025 – Salvador, Brazil (@robocup2025)

In the #RoboCup2025 @Home OPL Final, our robot performed very well. It opened two doors, removed trash, and closed a cabinet door. Overall, NimbRo came in second, next to team Tidyboy (Korea).
www.ais.uni-bonn.de/nimbro/@Home

[image or embed]

— Sven Behnke (@sven-behnke.bsky.social) 20 July 2025 at 18:04

View this post on Instagram

A post shared by RoboCupJunior Panamá (@robocupjrpa)

#RoboCup 2025 Competitions Day 3

Today was a tense day! The best robots in each category advanced to the final round that will take place tomorrow! Good luck to all the finalists #robocup2025 #robotics #ai pic.twitter.com/cC5PihLAge

— Asad Norouzi (@asadnorouzi) July 19, 2025

View this post on Instagram

A post shared by Nix-Lab (@nixlabmx)

View this post on Instagram

A post shared by RoboCup 2025 – Salvador, Brazil (@robocup2025)

ロボカップシンポジウム
Panel: The Future of RoboCup

大盛況です#Kyutech #KitaQ #HMA #RoboCup2025 #Salvador pic.twitter.com/eYCaFkiprb

— Hibikino-Musashi@Home (@HMA_wakamatsu) July 21, 2025

View this post on Instagram

A post shared by Juliana Vellanes (@julianavellanes)

We are in the finals!

We started with our own robot damaged during the transportation but now our super-team, including RoboCanes (U. of Miami), PUMAS (UNAM), @_erasers, and TIDbots (@jikei_tid), has got into the final at the #RoboCup at Home (DSPL) league!#RoboCup2025 pic.twitter.com/yFbCK7GEU6

— Luis Contreras (@TenshiTeacher) July 20, 2025

View this post on Instagram

A post shared by Fagner Ribeiro (@fagnerbribeiro)

View this post on Instagram

A post shared by Equipe de Robótica da UMI Tânia Leite Santos (@umibots)

View this post on Instagram

A post shared by RoboCup 2025 – Salvador, Brazil (@robocup2025)

View this post on Instagram

A post shared by autonOhm (@autonohm_official)

View this post on Instagram

A post shared by RoboCup 2025 – Salvador, Brazil (@robocup2025)

RoboCup2025 took place in Salvador, Brazil. The event saw around 3000 participants competing in the various leagues. Find out what the teams got up to during the first couple of days:

View this post on Instagram

A post shared by OBR (Olimpíada Brasileira de Robótica) (@olimpiadabrasileiraderobotica)

View this post on Instagram

A post shared by RoboCup 2025 – Salvador, Brazil (@robocup2025)

View this post on Instagram

A post shared by Matheus Moldes | MD – Design de interiores (@m3d_int)

View this post on Instagram

A post shared by TIGERs Mannheim (@tigersmannheim)

View this post on Instagram

A post shared by Reinaldo Bianchi (@bianchireinaldo)

View this post on Instagram

A post shared by Nix-Lab (@nixlabmx)

View this post on Instagram

A post shared by UP Robotics (@up_robotics_ags)

View this post on Instagram

A post shared by RoboCup 2025 – Salvador, Brazil (@robocup2025)

View this post on Instagram

A post shared by RoboCup 2025 – Salvador, Brazil (@robocup2025)

View this post on Instagram

A post shared by ER-Force Erlangen (@roboticserlangen)

View this post on Instagram

A post shared by RobôCIn (@robocinufpe)

View this post on Instagram

A post shared by @escherciliamoreira

View this post on Instagram

A post shared by UCIMU-SISTEMIPERPRODURRE (@ucimu.sistemiperprodurre)

View this post on Instagram

A post shared by Höhere Fachschule Technik Mittelland (@hftmag)

View this post on Instagram

A post shared by Brisbane Boys’ College Co-curricular (@bbccocurricular)

View this post on Instagram

A post shared by Flavio Dusse (@flaviodusse)

The annual RoboCup event, where teams gather from across the globe to take part in competitions across a number of leagues, will this year take place in Brazil, from 15-21 July. An important part of the week is RoboCupJunior, which is designed to introduce RoboCup to school children, and sees hundreds of kids taking part in a variety of challenges across different leagues. This year, the lead organizer for RoboCupJunior is Ana Patrícia Magalhães. We caught up with her to find out how the preparations are going, what to expect at this year’s competition, and how RoboCup inspires communities.

Could you tell us about RoboCupJunior and the plans you have for the competition this year?

RoboCup will take place from 15-21 July, in Salvador, Brazil. We expect to receive people from more than 40 countries, across the Junior and Major Leagues. We are preparing everything to accommodate all the students taking part in RoboCupJunior, who will participate in the Junior Leagues of Soccer, Rescue and OnStage. They are children and teenagers, so we have organized shuttles to take them from the hotels to the convention center. We’ve also prepared a handbook with recommendations about security, places they can visit, places to eat. The idea is to provide all the necessary support for them, because they are so young. We’re also organizing a welcome party for the Juniors so that they can experience a little bit of our culture. It will hopefully be a good experience for them.

The Juniors will be located on the first level of the mezzanine at the convention center. They will be separate from the Major Leagues, who will be on the ground floor. Of course, they’ll be able to visit the Major Leagues, and talk to the students and other competitors there, but it will be nice for them to have their own space. There will also be some parents and teachers with them, so we decided to use this special, dedicated space.

RoboCupJunior On Stage at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Do you have any idea of roughly how many teams will be taking part?

Yes, so we’ll have about 48 teams in the Soccer Leagues, 86 teams in the Rescue Leagues, and 27 in OnStage. That’s a lot of teams. Each team has about three or four students, and many of the parents, teachers and professors travel with them too. In total, we expect about 600 people to be associated with RoboCupJunior.

RoboCupJunior Soccer at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Have you got more RoboCupJunior participants from Brazil this year due to the location?

Yes, we have many teams from Brazil competing. I don’t know the exact number, but there are definitely more Brazilian teams this year, because it’s a lot cheaper and easier for them to travel here. When we have competitions in other countries, it’s expensive for them. For example, I have a team here in Salvador that qualified for the super regional event in the US and our team couldn’t go. They had qualified, but they couldn’t go because they didn’t have money to pay for the ticket. Now, it will be possible for all the Brazilian teams qualified to participate because it’s cheaper for them to come here. So it’s a big opportunity for development and to live the RoboCup experience. It’s very important for children and teenagers to share their research, meet people from other countries, and see what they are doing, and what research path they are following. They are very grateful for the opportunity to have their work tested against others. In a competition, it is possible to compare your research with others. So it’s different from conferences where you present a paper and show your work, but it’s not possible to compare and evaluate the results with other similar work. In a competition you have this opportunity. It’s a good way to get insights and improve your research.

RoboCupJunior Rescue at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

Your role at this RoboCup will be organizing RoboCupJunior. Are you also involved in the Major Leagues?

Yes, so my main role is organizing RoboCupJunior and I am also one of the chairs of the RoboCup Symposium. Besides, some teams from my laboratory are competing in the Major leagues. My team participates in the @Home league, but I haven’t had much time to help them recently, with all the preparations for RoboCup2025. Our laboratory also has teams from the 3d Simulation Soccer League, and the Flying Robots Demo. This will be the first time we’ll see a flying robot demo league at a RoboCup.

We’ll also have two junior teams from the Rescue Simulation League. They are very excited about taking part.

RoboCupJunior Rescue at RoboCup 2024. Photo: RoboCup/Bart van Overbeeke.

RoboCup was last held in Brazil in 2014, and I understand that there were quite a lot of new people that were inspired to join a team after that. Do you think the 2025 RoboCup will have the same effect and will inspire more people in Brazil to take part?

Yes, I hope so. The last one inspired many, many students. We could perceive the difference before and after RoboCup at that time, related to projects in schools. In 2014, RoboCup was held in João Pessoa, a city in the north east that is not as developed or populated as many other states in Brazil. It really improved the research in that place and the interest in robotics especially. After the 2014 RoboCup, we’ve had many projects submitted to the Brazilian RoboCup competition from that state every year. We believe that it was because of RoboCup being held there.

We hope that RoboCup2025 next month will have the same effect. We think it might have an even bigger impact, because there is more social media now and the news can spread a lot further. We are expecting many visitors. We will have a form where schools that want to visit can enroll on a guided visit of RoboCup. This will go live on the website next week, but we are already receiving many messages from schools asking how they can participate with their group. They are interested in the events, so we have high expectations.

We have been working on organizing RoboCup2025 for over a year, and there is still much to do. We are excited to receive everybody here, both for the competition and to see the city. We have a beautiful city on the coast, and some beautiful places to visit, so I recommend that people come and stay for some days after the competition to get to know our city.

About Ana Patrícia

Ana Patrícia F. Magalhães Mascarenhas received her PhD in Computer Science from the Federal University of Bahia (2016) and Master in Mechatronics from the Federal University of Bahia (2007). She is currently an adjunct professor at the State University of Bahia (UNEB) at the Information Systems course. She is a researcher and vice coordinator of the Center for Research in Computer Architecture, Intelligent Systems and Robotics (ACSO). Her current research focuses on service robotics and software engineering, especially related to the use of Artificial Intelligence (AI) in the software development process and in Model-Driven Development (DDM).

Claire chatted to Jane Pauline Ramos Ramirez from Delft University of Technology about drones that can move on land and in the air.

Jane Pauline Ramos Ramirez is a licensed engineer with a multidisciplinary background in bionics, mechanical, and aerospace engineering, and international research experience. Her life’s work is rooted in designing inclusive, socially accessible systems that work in synergy with nature and create meaningful impact in communities. As part of this mission, she has been developing nature-inspired drones that can move on both land and in the air — blending her appreciation for nature, design, and the mechanics of how things work.

By Kaja Šeruga

Just outside the historic German town of Goslar, a sprawling industrial complex receives an endless stream of discarded electronics. On arrival, this electronic waste is laboriously prepared for recycling. 

Electrocycling GmbH is one of the largest e-waste recycling facilities in Europe. Every year, it processes up to 80 000 tonnes of electronic waste, which comes in all shapes and forms.

Manual dismantling

Despite an impressive array of machinery, more than half of the site’s employees manually prepare the discarded items for recycling. They do this by sorting the incoming waste and removing batteries, which are a fire hazard and a major challenge in e-waste recycling.

“There are more and more devices, they are getting smaller, and they all contain lithium batteries, some of which are permanently installed, soldered or glued in place,” said Hannes Fröhlich, Electrocycling’s managing director. 

“It’s not a dream job, dismantling these appliances every day with hammers and pliers. I think we can do better.”

Some of these tedious tasks could be performed by robots. However, the problem is that every time there is a change in the product or the process, the hardware and software need to be restructured. This can be costly and time-consuming.

To address this issue, an EU-funded research initiative named ReconCycle has managed to automate the process by creating robots that can reconfigure themselves for different tasks. 

New territory for robotics

Researchers from Slovenia, Germany and Italy worked together on this issue at the Jožef Stefan Institute, Slovenia’s leading research facility, from 2020 to 2024.

The team developed adaptable AI-supported robots that are able to remove batteries from smoke detectors and radiator heat metres.

These two products can be found in most households and are replaced every five to eight years, creating large amounts of waste.

“The main challenge is that there are so many different versions of each device. Just think how many different remote controls there are,” said Dr Aleš Ude. He is head of the Department of Automatics, Biocybernetics and Robotics at the Jožef Stefan Institute and coordinates the ReconCycle research team. 

In industrial settings, robots are usually programmed for one specific task, repeating exactly the same series of movements in a predictable environment. 

Instead, the researchers set out to create a robot that can adapt to many different tasks, using state-of-the-art AI. 

“We wanted to expand robotics, introduce robots where there aren’t any yet,” Ude said.

A growing problem

Working with Electrocycling, Ude’s international research team created an adaptable robotic work cell. This is a workspace that consists of at least one robot, its tools and equipment, and its controller.

The novelty here is that this closed system autonomously adapts itself to various tasks, with the help of complex AI-driven software and modular hardware that can be quickly reconfigured. It also uses soft components like SoftHand, a human-like hand that can manipulate objects with great precision.

There are also safety features like collaborative robots and emergency stop buttons.

International collaboration was crucial in securing the right expertise, said Ude. 

“Robotics is very interdisciplinary, so it’s difficult to find the right partners in one country.” 

Thankfully, the new robots are arriving just at the right time, as the amount of e-waste produced every year continues to grow. Almost 5 million tonnes of e-waste are produced in the EU each year, amounting to about 11 kilograms per person. Less than 40% of that is recycled, the European Parliament has warned. 

Globally, around 62 million tonnes of e-waste were produced in 2022 alone, enough to fill 1.5 million 40-tonne trucks, according to UN data. Even more worryingly, the amount of e-waste is rising five times faster than the amount that is being recycled.

The EU is working to reduce e-waste through the Waste from Electrical and Electronic Equipment Directive, which sets the standards for collection and recycling. 

The work of Ude’s team is also aligned with the EU’s digital strategy, which encourages the use of AI in manufacturing to improve efficiency and help achieve climate neutrality by 2050.

Throwing away money

E-waste also has serious economic implications. An estimated €84 billion is lost each year when valuable metals like copper, iron and gold are discarded instead of being reused, according to the UN’s global e-waste monitor. 

At Electrocycling, 80% of the e-waste is recovered as raw materials, such as iron, zinc, gold, silver and palladium – some 35 materials in all.

“People need to understand that this is not just waste, but also raw materials that need to be recycled and kept in circulation, both for economic efficiency and a reduction of CO2,” said Fröhlich. 

New technology can make it even more efficient, and Fröhlich sees a lot of potential in it. 

“I was surprised by how far the technology and AI have already come,” he said. “They even recreated a human hand for the robot.”

Ude hopes to continue working with Electrocycling to improve e-waste solutions further. The hope is also that adaptable robots which can handle changing environments will have applications far beyond e-waste recycling. 

Given more time and development, these robots could even handle general housekeeping, or support carers in senior homes, said Ude. 

“Robotics could be of great help in such areas.”

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

Claire chatted to Will Kinghorn from Made Smarter about how to increase adoption of new tech by small manufacturers.

Will Kinghorn is an automation and robotics specialist for the Made Smarter Adoption Programme in the UK. With a background as a chartered manufacturing engineer in the aerospace industry, Will has extensive experience in developing and implementing automation and robotic solutions. He now works with smaller manufacturing companies, assessing their needs, identifying suitable technologies, and guiding them through the adoption process.  Last year he released a book called ‘Digital Transformation in Your Manufacturing Business – A Made Smarter Guide’.

By Kristýna Janovská and Pavel Surynek

Imagine if all of our cars could drive themselves – autonomous driving is becoming possible, but to what extent? To get a vehicle somewhere by itself may not seem so tricky if the route is clear and well defined, but what if there are more cars, each trying to get to a different place? And what if we add pedestrians, animals and other unaccounted for elements? This problem has recently been increasingly studied, and already used in scenarios such as warehouse logistics, where a group of robots move boxes in a warehouse, each with its own goal, but all moving while making sure not to collide and making their routes – paths – as short as possible. But how to formalize such a problem? The answer is MAPF – multi-agent path finding [Silver, 2005].

Multi-agent path finding describes a problem where we have a group of agents – robots, vehicles or even people – who are each trying to get from their starting positions to their goal positions all at once without ever colliding (being in the same position at the same time).

Typically, this problem has been solved on graphs. Graphs are structures that are able to simplify an environment using its focal points and interconnections between them. These points are called vertices and can represent, for example, coordinates. They are connected by edges, which connect neighbouring vertices and represent distances between them.

If however we are trying to solve a real-life scenario, we strive to get as close to simulating reality as possible. Therefore, discrete representation (using a finite number of vertices) may not suffice. But how to search an environment that is continuous, that is, one where there is basically an infinite amount of vertices connected by edges of infinitely small sizes?

This is where something called sampling-based algorithms comes into play. Algorithms such as RRT* [Karaman and Frazzoli, 2011], which we used in our work, randomly select (sample) coordinates in our coordinate space and use them as vertices. The more points that are sampled, the more accurate the representation of the environment is. These vertices are connected to that of their nearest neighbours which minimizes the length of the path from the starting point to the newly sampled point. The path is a sequence of vertices, measured as a sum of the lengths of edges between them.

Figure 1: Two examples of paths connecting starting positions (blue) and goal positions (green) of three agents. Once an obstacle is present, agents plan smooth curved paths around it, successfully avoiding both the obstacle and each other.

We can get a close to optimal path this way, though there is still one problem. Paths created this way are still somewhat bumpy, as the transition between different segments of a path is sharp. If a vehicle was to take this path, it would probably have to turn itself at once when it reaches the end of a segment, as some robotic vacuum cleaners do when moving around. This slows the vehicle or a robot down significantly. A way we can solve this is to take these paths and smooth them, so that the transitions are no longer sharp, but smooth curves. This way, robots or vehicles moving on them can smoothly travel without ever stopping or slowing down significantly when in need of a turn.

Our paper [Janovská and Surynek, 2024] proposed a method for multi-agent path finding in continuous environments, where agents move on sets of smooth paths without colliding. Our algorithm is inspired by the Conflict Based Search (CBS) [Sharon et al., 2014]. Our extension into a continuous space called Continuous-Environment Conflict-Based Search (CE-CBS) works on two levels:

Figure 2: Comparison of paths found with discrete CBS algorithm on a 2D grid (left) and CE-CBS paths in a continuous version of the same environment. Three agents move from blue starting points to green goal points. These experiments are performed in the Robotic Agents Laboratory at Faculty of Information Technology of the Czech Technical University in Prague.

Firstly, each agent searches for a path individually. This is done with the RRT* algorithm as mentioned above. The resulting path is then smoothed using B-spline curves, polynomial piecewise curves applied to vertices of the path. This removes sharp turns and makes the path easier to traverse for a physical agent.

Individual paths are then sent to the higher level of the algorithm, in which paths are compared and conflicts are found. Conflict arises if two agents (which are represented as rigid circular bodies) overlap at any given time. If so, constraints are created to forbid one of the agents from passing through the conflicting space at a time interval during which it was previously present in that space. Both options which constrain one of the agents are tried – a tree of possible constraint settings and their solutions is constructed and expanded upon with each conflict found. When a new constraint is added, this information passes to all agents it concerns and their paths are re-planned so that they avoid the constrained time and space. Then the paths are checked again for validity, and this repeats until a conflict-free solution, which aims to be as short as possible is found.

This way, agents can effectively move without losing speed while turning and without colliding with each other. Although there are environments such as narrow hallways where slowing down or even stopping may be necessary for agents to safely pass, CE-CBS finds solutions in most environments.

This research is supported by the Czech Science Foundation, 22-31346S.

You can read our paper here.

References

• Janovská, K. and Surynek, P. (2024). Multi-agent Path Finding in Continuous Environment, CoRR.
• Sharon, G., Stern, R., Felner, A., and Sturtevant, N. R. (2014). Conflict-based search for optimal multi-agent pathfinding, Artificial Intelligence.
• Karaman, S. and Frazzoli, E. (2011). Sampling-based algorithms for optimal motion planning, CoRR.
• Piegl, L. and Tiller, W. (1996). The NURBS Book, Springer-Verlag, New York, USA, second edition.
• Silver, D. (2005). Cooperative pathfinding, Proceedings of the First Artificial Intelligence and Interactive Digital Entertainment Conference, Marina del Rey, California, USA.

There are two pieces of research we’d like to ask you about today. Could we start with the GenWarp paper? Could you outline the problem that you were focused on in this work?

The problem we aimed to solve is called single-shot novel view synthesis, which is where you have one image and want to create another image of the same scene from a different camera angle. There has been a lot of work in this space, but a major challenge remains: when an image angle changes substantially, the image quality degrades significantly. We wanted to be able to generate a new image based on a single given image, as well as improve the quality, even in very challenging angle change settings.

How did you go about solving this problem – what was your methodology?

The existing works in this space tend to take advantage of monocular depth estimation, which means only a single image is used to estimate depth. This depth information enables us to change the angle and change the image according to that angle – we call it “warp.” Of course, there will be some occluded parts in the image, and there will be information missing from the original image on how to create the image from a new angle. Therefore, there is always a second phase where another module can interpolate the occluded region. Because of these two phases, in the existing work in this area, geometrical errors introduced in warping cannot be compensated for in the interpolation phase.

We solve this problem by fusing everything together. We don’t go for a two-phase approach, but do it all at once in a single diffusion model. To preserve the semantic meaning of the image, we created another neural network that can extract the semantic information from a given image as well as monocular depth information. We inject it using a cross-attention mechanism, into the main base diffusion model. Since the warping and interpolation were done in one model, and the occluded part can be reconstructed very well together with the semantic information injected from outside, we saw the overall quality improved. We saw improvements in image quality both subjectively and objectively, using metrics such as FID and PSNR.

Can people see some of the images created using GenWarp?

Yes, we actually have a demo, which consists of two parts. One shows the original image and the other shows the warped images from different angles.

Moving on to the PaGoDA paper, here you were addressing the high computational cost of diffusion models? How did you go about addressing that problem?

Diffusion models are very popular, but it’s well-known that they are very costly for training and inference. We address this issue by proposing PaGoDA, our model which addresses both training efficiency and inference efficiency.

It’s easy to talk about inference efficiency, which directly connects to the speed of generation. Diffusion usually takes a lot of iterative steps towards the final generated output – our goal was to skip these steps so that we could quickly generate an image in just one step. People call it “one-step generation” or “one-step diffusion.” It doesn’t always have to be one step; it could be two or three steps, for example, “few-step diffusion”. Basically, the target is to solve the bottleneck of diffusion, which is a time-consuming, multi-step iterative generation method.

In diffusion models, generating an output is typically a slow process, requiring many iterative steps to produce the final result. A key trend in advancing these models is training a “student model” that distills knowledge from a pre-trained diffusion model. This allows for faster generation—sometimes producing an image in just one step. These are often referred to as distilled diffusion models. Distillation means that, given a teacher (a diffusion model), we use this information to train another one-step efficient model. We call it distillation because we can distill the information from the original model, which has vast knowledge about generating good images.

However, both classic diffusion models and their distilled counterparts are usually tied to a fixed image resolution. This means that if we want a higher-resolution distilled diffusion model capable of one-step generation, we would need to retrain the diffusion model and then distill it again at the desired resolution.

This makes the entire pipeline of training and generation quite tedious. Each time a higher resolution is needed, we have to retrain the diffusion model from scratch and go through the distillation process again, adding significant complexity and time to the workflow.

The uniqueness of PaGoDA is that we train across different resolution models in one system, which allows it to achieve one-step generation, making the workflow much more efficient.

For example, if we want to distill a model for images of 128×128, we can do that. But if we want to do it for another scale, 256×256 let’s say, then we should have the teacher train on 256×256. If we want to extend it even more for higher resolutions, then we need to do this multiple times. This can be very costly, so to avoid this, we use the idea of progressive growing training, which has already been studied in the area of generative adversarial networks (GANs), but not so much in the diffusion space. The idea is, given the teacher diffusion model trained on 64×64, we can distill information and train a one-step model for any resolution. For many resolution cases we can get a state-of-the-art performance using PaGoDA.

Could you give a rough idea of the difference in computational cost between your method and standard diffusion models. What kind of saving do you make?

The idea is very simple – we just skip the iterative steps. It is highly dependent on the diffusion model you use, but a typical standard diffusion model in the past historically used about 1000 steps. And now, modern, well-optimized diffusion models require 79 steps. With our model that goes down to one step, we are looking at it about 80 times faster, in theory. Of course, it all depends on how you implement the system, and if there’s a parallelization mechanism on chips, people can exploit it.

Is there anything else you would like to add about either of the projects?

Ultimately, we want to achieve real-time generation, and not just have this generation be limited to images. Real-time sound generation is an area that we are looking at.

Also, as you can see in the animation demo of GenWarp, the images change rapidly, making it look like an animation. However, the demo was created with many images generated with costly diffusion models offline. If we could achieve high-speed generation, let’s say with PaGoDA, then theoretically, we could create images from any angle on the fly.

Find out more:

• GenWarp: Single Image to Novel Views with Semantic-Preserving Generative Warping, Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, Yuki Mitsufuji.
• GenWarp demo
• PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher, Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Yuhta Takida, Naoki Murata, Toshimitsu Uesaka, Yuki Mitsufuji, Stefano Ermon.

About Yuki Mitsufuji

Yuki Mitsufuji is a Lead Research Scientist at Sony AI. In addition to his role at Sony AI, he is a Distinguished Engineer for Sony Group Corporation and the Head of Creative AI Lab for Sony R&D. Yuki holds a PhD in Information Science & Technology from the University of Tokyo. His groundbreaking work has made him a pioneer in foundational music and sound work, such as sound separation and other generative models that can be applied to music, sound, and other modalities.

Claire chatted to Robert Siddall from the University of Surrey about novel robot designs inspired by the way real animals move.

Robert Siddall is an aerospace engineer with an enthusiasm for unconventional robotics. He is interested in understanding animal locomotion for the benefit of synthetic locomotion, particularly flight. Before becoming a Lecturer at the University of Surrey, he worked at the Max Planck Institute for Intelligent Systems in Stuttgart, Germany, where he studied the arboreal acrobatics of rainforest-dwelling reptiles. His work focuses on the design of novel robots that can tackle important environmental problems.

Claire chatted to Gianmarco Pisanelli from the the University of Sheffield Advanced Manufacturing Research Centre about how to promote the safe and intuitive use of robots in manufacturing.

Gianmarco Pisanelli specialises in early-stage technology readiness level (TRL) research, with a focus on collaborative workspaces, multi-robot systems, and robotic simulations. He possesses extensive experience in industrial robotics, including PLC programming and the development of intuitive robotic interfaces such as natural speech and augmented reality. As the lead developer for the Robot Operating System (ROS) at the University of Sheffield’s Advanced Manufacturing Research Centre (AMRC), he serves as the primary point of contact for SMEs and major partners.

Claire chatted to Isabella Fiorello from the University of Freiburg about plant-inspired robots made from living materials.

Isabella Fiorello is a Junior Group Leader and Principal Investigator of the Bioinspired Plant-hybrid Materials group at the University of Freiburg in Germany. She has a Master’s Degree in Industrial Biotechnology from the University of Turin in Italy and a PhD in Biorobotics from Scuola Superiore Sant’Anna in Italy. Her research focusses on the development of biologically-inspired microfabricated living materials able to precisely interact with complex unstructured surfaces for applications in precision agriculture, smart fabrics, space, and soft robotics.

Claire chatted to Mini Rai from Orbit Rise about orbital and planetary robots.

Mini Rai is the founding Director of Orbit Rise Ltd and an honorary Professor at the University of Lincoln. Previously, she was the Global Chair in Robotic Engineering at the University of Lincoln and an Associate Professor at the Surrey Space Centre. Mini has over 27 years of research and innovation experience in Space Engineering and Technology. With deep-rooted expertise in robotics, automation, control and systems engineering, she has led a large and diverse portfolio of national and international programmes on space robotic missions, spanning orbital and planetary robotics.

The 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024) was held from 14-18 October in Abu Dhabi, UAE. We take a look at what the participants got up to.

Join us today for the Brain over Brawn (BoB) Workshop on Label Efficient Learning Paradigms for Autonomy at Scale #IROS2024
The program is exciting:
– speakers: @ftm_guney @lucacarlone1 @peymanm @mpopovic514
– panelists: @_ayoungk @stachnis @maththrills https://t.co/t9RMqKFT2U pic.twitter.com/NGUyjgqX8Q

— Andrei Bursuc (@abursuc) October 14, 2024

#IROS2024 attendees, come and see our presentation on wednesday morning!

WePI2T09.07 | Wed 16 Oct 2024 | Room 9
[1] Teaser Session (Navigation I):
09:00 – 10:00
[2] Interactive session: 10:00 – 12:00 in the poster area near the Room 9. https://t.co/Bs3M5gDzaU pic.twitter.com/HaZBmuUqK1

— MIZUHO AOKI (@mizuhoaoki1998) October 14, 2024

In #IROS2024 Workshop, "Collecting, Managing and Utilizing Data through Embodied Robots" (https://t.co/d81DRq1JGk) was successfully held!
Thank you so much for all of speakers,
presenters, audiences!@namikosaito_ @tetsuyaogata1 @hshige_86 @rsk_tmr @funabashihand @RMalsada3 https://t.co/ZA9T8qehDH pic.twitter.com/sdXBWhe79Q

— Tamon Miyake (@T_Miyake3) October 14, 2024

What about 12 talks in parallel in the same room?
Personal headphones for everyone make it possible.
@IEEEorg #IROS2024 Abu Dhabi is trying this quite a peculiar setup which is working well and keep everyone in a single open space. pic.twitter.com/PDXCdyKrsT

— Giovanni Reina (@joefresna) October 14, 2024

Thanks @BotJunkie from @IEEESpectrum and Lucy from @AIhuborg / @Robohub for the #IROS2024 scicomm training! Great content from 'why to communicate', to 'how to write a story, find images, and avoid hype', and finally 'engaging with media'. Great it was available online @ieeeras! pic.twitter.com/aFPbO5WTzH

— Sabine Hauert (@sabinehauert) October 15, 2024

Live demo from plenary! Remote operation from Abu Dhabi to Stanford. #IROS2024 pic.twitter.com/4YsGicbZ4M

— Ayoung (@_ayoungk) October 15, 2024

Early morning talk in Abu Dhabi = 1am talk in Boston

Sad to miss @IEEEorg #IROS2024 in person but glad I still had a chance to share the latest updates on biohybrid robotics from our lab @MITMechE!

Now for some caffeine pic.twitter.com/Ph8RWOdZa9

— Ritu Raman (@DrRituRaman) October 15, 2024

Not all heroes wear capes, sometimes they carry external GPU for robot #IROS2024 pic.twitter.com/rZ2CUHGgpG

— Michael Cho – Rbt/Acc (@micoolcho) October 16, 2024

OMRON SINIC X at #IROS2024! (4/7)
Title: Robot Swarm Control Based on Smoothed Particle Hydrodynamics for Obstacle-Unaware Navigation
Date: Thu, Oct 17, 11:30-11:45 | Room14
Oral Presentations: Paper ThBT14.3
Learn more: https://t.co/lFdkNv6F6H
YouTube: https://t.co/3wLKbvqCAg

— OMRON SINIC X (OSX) (@omron_sinicx) October 17, 2024

Some exciting work is being showcased at the IEEE/RSJ International Conference on Intelligent Robots and Systems. In his research, Javier proposes methods for reconfigurable robots to navigate safely. #IROS2024 #ReconfigurableRobots #AbuDhabi #UAE pic.twitter.com/Yls06W0No7

— IEEE RAS (@ieeeras) October 17, 2024

Day 3 at #IROS2024 is off to a great start!
Yesterday was action-packed with interactive demos and insightful discussions. If you haven't visited yet, be sure to stop by booth 49—we're excited to meet you!
Check out some highlights from yesterday!#HatsuMuv #HatsuHand pic.twitter.com/jpV8bUvDNv

— HatsuMuv Official (@HatsuMuv) October 17, 2024

We had a wonderful time at the Workshop on Interactive Robots and AI for Healthcare at #IROS2024 at ADNEC Centre Abu Dhabi. We enjoyed engaging in discussions with friends and new collaborators.https://t.co/vH2xIsiDXQ@BristolRobotLab @ieeeras @UWEBristol pic.twitter.com/eOB3ONfYL1

— Carlos Cifuentes (@cacifuentesg) October 17, 2024