You may have seen a lot of press around a dashcam video of a car accident in the Netherlands. It shows a Tesla in AutoPilot hitting the brakes around 1.4 seconds before a red car crashes hard into a black SUV that isn’t visible from the viewpoint of the dashcam. Many press have reported that the Tesla predicted that the two cars would hit, and because of the imminent accident, it hit the brakes to protect its occupants.

The accident is brutal but apparently nobody was hurt.

https://www.youtube.com/watch?v=lqqN5iRrAiM

The press speculation is incorrect. It got some fuel because Elon Musk himself retweeted the report linked to, but Telsa has in fact confirmed the alternate and more probable story which does not involve any prediction of the future accident. In fact, the red car plays little to no role in what took place.

Tesla’s autopilot uses radar as a key sensor. One great thing about radar is that it tells you how fast every radar target is going, as well as how far away it is. Radar for cars doesn’t tell you very accurately where the target is (roughly it can tell you what lane a target is in.) Radar beams bounce off many things, including the road. That means a radar beam can bounce off the road under a car that is in front of you, and then hit a car in front of it, even if you can’t see the car. Because the radar tells you “I see something in your lane 40m ahead going 20mph and something else 30m ahead going 60mph” you know it’s two different things.

The Tesla radar saw just that — the black SUV was hitting the brakes (possibly for a dirt patch that appears to show on the video) and the red car wasn’t. Regardless of the red car being there, the autopilot knew that if another car ahead was braking hard, it should also brake hard, and it did. Yes, it’s possible that it could also calculate that the red car, if it keeps going, will hit the black car, but that’s not entirely relevant — it’s clear that the Tesla should stop, regardless of what the red car is going to do. Tesla reported in their blog about how they were doing more with the radar, including tracking hidden vehicles with it. The ability of automotive radar to do this has been known for some time, and I have always presumed most teams have taken advantage of it. You don’t always get returns from hidden cars, but it’s worth using them if you do.

In the future, we will see robocar systems predicting accidents, but I am not aware of this being announced by any team. All robocars are tracking all objects ahead of them, for position and velocity, and they are extrapolating their velocity and predicting where they will go. Those predictions would also include detecting that vehicles might hit (if they continue their current course) and also if they could not avoid hitting at a certain point. If an imminent accident is predicted, it would make sense to know that and also react to it in advance. A car might even be able to predict a bit of what will happen after the accident, though that is chaotic.

A system like that would outperform the autopilot or any automatic emergency braking system. Presently, those systems largely track objects in their lane. They don’t brake because cars are stopped in adjacent lanes, because that would mean they could not work in traffic jams or carpool lanes when there are lanes at different speeds, and they could not deal with stalled cars on the side of the road.

However, if you saw what the Tesla saw from the lane to the right, it would still be a very smart thing to brake. Tesla has not commented on this, but I presume its system would not have braked if it had been in that lane, at least not braked before the accident. It might brake because other cars like the red car immediately moved into the right lane.

If you liked this post on robocars, you’ll also enjoy these articles:

• What if the city ran Waze and you had to obey it? Could this cure congestion?
• What are the right disability rules for robotaxis?
• Google car is now Waymo
• Comma.ai cancels comma-one add-on box after threats from NHTSA
• Robocar news: Comma One goes open source, creating simulations for robocars in New Zealand earthquakes

See all the latest robotics news on Robohub, or sign up for our weekly newsletter

Source: commaai

Comma.ai, the brash startup attempting to make a self-driving system entirely from a neural network has announced it will cancel the “comma one” add-on box it has planned to sell to owners of certain Honda vehicles. The box stuck on the rear-view mirror and used the car’s own bus commands to provide an autopilot similar to those offered by car makers, with lane-keeping and adaptive cruise control.

Of particular importance is the letter from NHTSA to comma.ai which I suggest you read. This letter creates several big issues:

1. There are many elements of this letter that would also apply to Tesla and other automakers which have built supervised autopilot functions.
2. Of particular interest is the paragraph that says: “it is insufficient to assert, as you do, that the product does not remove any of the driver’s responsibilities” and “there is a high likelihood that some drivers will use your product in a manner that exceeds its intended purpose.” That must be very scary for Tesla.
3. I noted before that the new NHTSA regulations appear to forbid the use of “black box” neural network approaches to the car’s path planning and decision making. I wondered if this made illegal the approach being done by Comma, NVIDIA and many other labs and players. This may suggest that.
4. We now have a taste of the new regulatory regime and had it existed before systems like Tesla’s Autopilot, Mercedes Traffic Jam Assist, and Cruise’s original aftermarket autopilot would never have been able to get off the ground.
5. George Hotz of comma declares, “Would much rather spend my life building amazing tech than dealing with regulators and lawyers. It isn’t worth it. The comma one is cancelled. comma.ai will be exploring other products and markets. Hello from Shenzhen, China.”

To be clear, comma is a tiny company taking a radical approach, so it is not a given that what NHTSA has applied to them would have been or will be unanswerable by the big guys. Because Tesla’s Autopilot is not a pure machine learning system, they can answer many of the questions in the NHTSA letter that comma can’t. They can do much more extensive testing that a tiny startup can’t. But even so, a letter like this sends a huge chill through the industry.

It should also be noted that in Comma’s photos the box replaced the rear-view mirror, and NHTSA had reason to ask about that.

Source: commaai

George’s declaration that he’s in Shenzen gives us the first sign of the new regulatory regime pushing innovation away from the United States and California. I will presume the regulators will say, “We only want to scare away dangerous innovation” but the hard truth is that is a very difficult thing to judge. All innovation in this space is going to be a bit dangerous. It’s all there trying to take the car — the 2nd most dangerous legal consumer product — and make it safer, but it starts from a place of danger. We are not going to get to safety without taking risks along the way.

I sometimes ask, “Why do we let 16-year-olds drive?” They are clearly a major danger to themselves and others. Driver testing is grossly inadequate. They are not adults so they don’t have the legal rights of adults. We let them drive because they are going to start out dangerous and then get better. It is the only practical way for them to get better, and we all went through it. Today’s early companies are teenagers. They are going to take risks. But this is the fastest and only practical way to let them get better and save millions.

“…some drivers will use your product in a manner that exceeds its intended purpose”

This sentence, though in the cover letter and not the actual legal demand, looks at the question asked so much after the Tesla fatal crash. The question which caused Consumer Reports to ask Tesla to turn off the feature. The question which caused MobilEye, they say, to sever their relationship with Tesla.

The paradox of the Autopilot is this: The better it gets, the more likely it is to make drivers over-depend on it. The more likely they will get complacent and look away from the road. And thus, the more likely you will see a horrible crash like the Tesla fatality. How do you deal with a system which adds more danger the better you make it? Customers don’t want annoying countermeasures. This may be another reason that “Level 2,” as I wrote yesterday is not really a meaningful thing.

NHTSA has put a line in the sand. It is no longer going to be enough to say that drivers are told to still pay attention.

Black box

Comma is not the only company trying to build a system with pure neural networks doing the actual steering decisions (known as “path planning”.) NVIDIA’s teams have been actively working on this, as have several others. They plan to make commentary to NHTSA about these elements of the regulations, which should not be forbidding this approach until we know it to be dangerous.

It is challenging. It’s hard to do QA on neural networks. You can examine any single state of them, but not really understand them. You can fix the errors they make, but not know how you fixed it or whether your fix is going to work in other cases. On their own that sounds too scary, but the problem is they are outperforming other algorithms at many of the problems they are being applied to.

If you have two systems:

1. A conventionally coded system which you fully understand, that has one safety incident per 100,000 miles,
2. A black box machine learning system that has one safety incident per 150,000 miles, but you have minimal understanding of.

Which is the one that is better to put on the road? It’s not a no-brainer.

Scaring people away

George Hotz is about as brash as they come, so some might discount his bluster about moving his projects to China. But this is a real thing. When regulations come, they tend to favour the big players. The big players have large QA teams and compliance departments. They can afford them, and they move conservatively because of them. Big players help write the regulations to favour their way of doing things — sometimes selfishly but sometimes just because it is the only thing they know.

The USA and California led the way in robocars in part because it was unregulated. In the USA, everything is permitted unless it was explicitly forbidden and nobody thought to write “no robots” in the laws. Progress in other countries where everything is forbidden unless it is permitted was much slower. The USA is moving in the wrong direction.

Countries like China are another animal altogether. There, everything is forbidden unless a smart autocrat allows it but once the smart (or stupid) ruling official makes it happen, the other red tape can vanish. It is not the best system overall, but it can win at things they put their minds to.

Tesla

Has Tesla received such a letter? What would they have done if they received it in their early days? While Tesla and Elon Musk love the Autopilot project, the hard truth is Tesla’s biggest problem is going to be making enough Model 3s to meet demand and doing it profitably. Taking out Autopilot isn’t going to cut their sales at all — they have too many sales and not enough cars. Only their forward vision and guts will keep them going if NHTSA does the same thing to them. I do believe Tesla can answer questions like those in the demand letter more easily than comma.ai. But not trivially, and they almost surely will need to bump the countermeasures.

Tesla Model S autopilot-software. Source: Tesla

Tesla’s spat with MobilEye reached a new pitch this week with Tesla announcing a new release of their autopilot and plans. As reported earlier, MobilEye announced during the summer that they would not be supplying the new and better versions of their EyeQ system to Tesla. Since that system was/is central to the operation of the Tesla Autopilot, they may have been surprised that MBLY stock took a big hit after the announcement (though it recovered for a while, and is now back down) and TSLA did not.

Statements and documents now show a nastier battle with MobilEye saying they were worried about Tesla using their tool in an unsafe way, invoking all the debate about the fatality and other crashes allegedly caused by people who were lulled into not bothering to supervise the Autopilot. Tesla says that, instead, they have been developing their own advanced vision tools and that MobilEye was afraid and told Tesla if they wanted more EyeQ chips, they would need to halt the competing project and commit to ME. That’s a nasty spat.

Tesla’s own efforts represent a threat to MobilEye from the growing revolution in neural network pattern matchers. Computer vision is going through a big revolution. MobilEye is a big player in that revolution because their ASICs do both standard machine vision functions and can do neural networks. An ASIC will beat a general purpose processor when it comes to cost, speed, and power, but only if the ASIC’s abilities were designed to solve those particular problems. Since it takes years to bring an ASIC to production, you have to aim right. MobilEye aimed pretty well, but at the same time, lots of research out there is trying to aim even better or do things with more general purpose chips, like GPUs. Soon we will see ASICs aimed directly at neural network computations. To solve the problem with neural networks, you need the computing horsepower, and you need well designed deep network architectures, plus, you need the right training data and lots of it. Tesla and ME both are gaining lots of training data. Many companies, including Nvidia, Intel and others are working on the hardware for neural networks. Most people would point to Google as the company with the best skills in architecting the networks, though there are many doing interesting work there. (Google’s DeepMind built the tools that beat humans at the seemingly impossible game of Go, for example.) It’s definitely a competitive race.

Using radar

While Tesla works on their vision systems, they also announced a plan to make much more use of radar. That’s an interesting plan. Radar has been the poor 3rd-class sensor of the robocar, after LIDAR and vision. Everybody uses it — you would be crazy not to unless you need to be very low cost. Radar sees further than the other systems, and it tells you immediately how fast any radar target you see is moving relative to you. It sees through fog and other weather, and it can even see under and around big cars in front of you as it bounces off the road and other objects. It’s really good at licence plates, as well.

What radar doesn’t have is high resolution. Today’s automotive radars have gotten good enough to tell you what lane an object like another car is in, but they are not designed to have any vertical resolution — you will get radar returns from a stalled car ahead of you on the road and a sign above that lane, and not be sure of the difference. You need your car to avoid a stalled car in your lane, but you can’t have a car that hits the brakes every time it sees a road sign or bridge!

mage source: EDN Network

Real world radar is messy. Your antennas send out and receive from a very broad cone with potential signals from other directions and from side lobes. Reflections are coming from vehicles and road users but also from the ground, hills, trees, fences, signs, bushes and bridges. It’s work to get reliable information from it. Early automotive radars found the best solution was to use the doppler speed information, and discard all returns from anything that wasn’t moving towards or away from you — including stalled cars and cross traffic.

One thing that can help (imperfectly) is a map. You can know where the bridges and signs are so you don’t brake for them. Now you can brake for the stalled cars and the cross traffic the Tesla failed to see. You still have an issue with a stalled car under a bridge or sign, but you’re doing a lot better.

There’s a lot of room for improvement in radar and I will presume — Tesla has not said — that Tesla plans to work on this. The automotive radars everybody buys (from companies like Bosch) were made for the ADAS market — adaptive cruise control, emergency braking etc. It is possible to design new radars with more resolution (particularly in the vertical) and other approaches. You can also try for more resolution, particularly by splitting the transmitter and receiver to produce a synthetic larger aperture. You can go into different bands and get more bandwidth and get more resolution in general. You can play more software tricks, and particularly, you can learn by examining not just single radar returns, but rather the pattern of returns over time. (After all, humans don’t navigate from still frames, we depend on our visual system’s deep evolved ability to use motion and other clues to understand the world.)

The neural networks are making strides here. For example, while pedestrians produce basic radar returns, it turns out that their walking stride has a particular pattern of changes that can be identified by neural networks. People are doing research now on how examining the moving and dynamic pattern of radar returns can help you get more resolution and also identify shapes and motion patterns of objects and figure out what they are.

I will also speculate that it might be possible to return to a successor of the “sweeped” radars of old, the ones we are used to seeing in old war movies. Modern car radars don’t scan like that, but I have to wonder if with new techniques, like phased arrays to steer virtual beams (already the norm in military radar) and modern high-speed electronics, that we might produce radars that get a better sense of where their target is. We’re also getting better at sensor fusion — identifying a radar target in an image or LIDAR return to help learn more about it.

The one best way to improve radar resolution would be to use more bandwidth. There have been experiments in using ultrawideband signals in the very high frequencies which may offer promise. As the name suggests, UWB uses a very wide band, and it distributes its energy over that very wide band, which means it doesn’t put too much energy into any one band and has less chance of interfering in those bands. It’s also possible that the FCC, seeing the tremendous public value that reliable robocars offer, might consider opening up more spectrum for use in radar applications using modern techniques and thus increase the resolution.

In other words, Tesla is wise to work on getting more from radar. With the loss of all MobilEye’s vision tools, they will have to work hard to duplicate and surpass that. For now, Tesla is committed to using parts that are for sale for existing production cars, costing hundreds of dollars. That has taken LIDAR “off their radar” even though almost all research teams depend on LIDAR and expect LIDAR to be cheap in a couple of years. (Including the LIDAR from Quanergy, a company I advise.)

Comma announces a $1,000 autopilot box

I wrote earlier about comma.ai and their efforts to drive with just vision, radar and neural networks. They now plan to offer a box for $1,000 to give you some basic autopilot functionality as an add-on.

To do this, they are working with only some specific car models, namely some Honda vehicles that already have advanced ADAS in them. Using the car’s internal bus, they can talk to the sensors in these cars (such as the radar and possibly the camera) and also send control signals to actuate the steering, brakes and throttle. Then their neural networks can take the sensor information, and output the steering and speed commands to keep you in the lane. (Details are scant so I don’t know if the Comma One box uses its own camera or depends on access to the car’s.)

When I rode in Comma’s prototype it certainly wasn’t up to the level of the Tesla autopilot or some others, but it has been several months so I can’t judge it now. Like the Tesla autopilot, the Comma will not be safe enough to drive the car on its own, and you will need to supervise and be ready to intervene at any time. If you get complacent, as some Tesla drivers have, you could get injured or killed. I have yet to learn what measures Comma will take to make sure people keep their eyes on the road.

Generally, I feel that autopilots are not very exciting products when you have to watch them all the time — as you do — and that bolt-on products are not particularly exciting. Cruise’s initial plan (after they abandoned valet parking) was a bolt-on autopilot, but they soon switched to trying to build a real vehicle, and that got them the huge $700M sale to General Motors.

But for Comma, there is a worthwhile angle. Users of this bolt-on box will be helping to provide training data to improve their systems. In fact, they will be paying for the privilege of testing the system and training it. Something that companies like Google did the old-fashioned way, paying a staff of professionals to drive the cars and gather data. For a tiny, young startup it’s a worthwhile approach.

If you liked this article, you may also want to read:

• Robocar recap: Uber buys Otto, folks leave Google, Ford goes big, Tesla drops MobilEye
• Will robocars be heaven or hell for our cities?
• Robocar platooning, or just carpool?
• Understanding the massive gulf between the Tesla Autopilot and a real robocar, in light of the crash

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Tesla’s autopilot isn’t even particularly new. Several car makers have had similar products in their labs for several years, and some have released it to the public, at first in a “traffic jam assist” mode, but reportedly in full highway cruise mode outside the USA. The first companies to announce it were Cadillac with the “Super Cruise” and VW’s “Temporary Autopilot” — but they delayed that until much later.

Remarkably, Honda showed off a car ten years ago doing this sort of basic autopilot (without lane change) and sold only in the UK. They decided to discontinue the project, however. That this was actually promoted as an active product ten years ago will give you some clue as to how different this was from the bigger efforts.

Cruise products like these require constant human supervision. With regular cruise control, you could take your feet off the pedals, but might have to intervene fairly often either by using the speed adjust buttons or full control. Interventions could be several times a minute. Later came “Adaptive Cruise Control”, which still required you to steer and fully supervise, but would rarely require intervention on the pedals while driving on the highway — a few times an hour might be acceptable.

The new autopilot systems allow you to take your hands off the wheel but demand full attention. Users report needing to intervene rarely on some highways, but frequently on other roads. Once again, if you only need to intervene once an hour, the product could make your drive more relaxing.

Now consider a car that drives without supervision …

Human drivers have minor accidents about every 2,500 to 6,000 hours, depending on what figures you are using — that would be about once every 10 to 20 years of driving. A fatal accident takes place every 2,000,000 hours of driving — about once every 10,000 years for the typical driver, thankfully a much longer span than a person’s lifetime.

If a full robocar needs human intervention, logic tells you that it’s going to have an accident because there is nobody there to intervene. Just like with human drivers, most of the errors that would cause an accident are minor: running off the road, fender benders. Not every mistake that could cause a crash or a fatality causes one. Indeed, humans make mistakes that might cause a fatality far more often than every 2,000,000 hours, because we “get away” with many of them.

Even so, the difference is staggering. A cruise autopilot (such as Tesla’s) is a workable product if you have to correct it a few times an hour, whereas a full robocar product is only workable if you need to correct it only after decades or even lifetimes of driving. This is not a difference of degree, it is a difference of kind. It is why there is probably not an evolutionary path from the cruise/autopilot systems based on existing ADAS technologies to a real robocar. Doing many thousands times better will not be done by incremental improvement. It almost surely requires a radically different approach, and probably very different sensors.

To top it all off, a full robocar doesn’t just need to be great at avoiding accidents. If it’s running unmanned, with no human to help it at all, it needs a lot of other features and capabilities too.

The mistaken belief in an evolutionary path also explains why some people imagine robocars are many decades away. If you wanted evolutionary approaches to take you to 100,000x better, you would expect to wait a long time. When an entirely different approach is required, what you learn from the old approach doesn’t help you predict how the other approaches — including unknown ones — will do.

It does teach you something, though. By simply being on the road, Tesla will encounter all sorts of interesting situations its developers weren’t expecting, and they will use this data to train new generations of software that do better. They will learn things that help them make the revolutionary unmanned product they hope to build in the 2020s. This is a good thing.

Google and others have also been out learning that, and soon more teams will.

This post originally appeared on Robocars.com.

If you liked this article, you may also be interested in:

• The two cultures of robocars
• Might the first, supervised robocars be… well… boring?
• Issues in regulating robocars, and the case for a light hand
• Future driving: Business models for robocars

See all the latest robotics news on Robohub, or sign up for our weekly newsletter.

Airware also just announced the launch of its new generic autopilot, control station, and operating system for unmanned fixed-wing and multirotor aircraft. The two red Airware boxes comply with government and insurance requirements and will work with most drone builders’ devices. Airware hopes to become the standard for commercial drones by enabling customers to select drones and then mix and match hardware and software components to create unique devices for different jobs; they provide hardware, software, and cloud services that enable reliable and scalable enterprise drone operations.

With Airware focusing on their autopilot boxes and software systems, global drone makers and startups can quickly build safe and reliable drones, pop in the Airware boxes, and supply their special application-specific software. Thus drone manufacturers can tailor and focus their drones to any commercial application by connecting aircraft, sensors, payloads, and application-specific software, and not have to develop extensive piloting and navigation software themselves.

COLLISION AVOIDANCE AND TRANSPONDERS

Other companies are also developing autopilot systems. 3D Robotics offers their open source DroneKit; ArduPilot has an autopilot as does MicroPilot. A couple of defense contractors also have autopilot products and some are working on collision avoidance systems. Mitre and Aerialtronics are two such companies. But Airware’s proprietary software is more commercial-based and is already being field tested for mining surveys, precision agriculture, industrial inspections and forestry. And it is likely that they will add collision avoidance and a transponder once those systems get miniaturized and approved by the authorities.

“Airware provides the market with a full spectrum of products that bring real added value to our systems,” said Christian F. Viguie, chairman and CEO of Delta Drone, “they are building a powerful ecosystem that unites this growing market.”

General Electric announced that they will become Airware’s first large enterprise customer. Alex Tepper, managing director at GE Ventures said “We are currently developing drone solutions for our customers … drones have the ability to reduce downtime, increase safety, and provide more reliable operations for our customers. We believe that Airware is going to be a key partner in helping us deliver these solutions.”

There is much interest amongst the VC community to be invested in the emerging and exponentially growing drone industry as can be seen in these two articles:

• Consumer and B2B drone making is big business
• Commercial drones market to grow at 109% CAGR

Many research reports forecast exponential growth such as the one referenced above. Thus it is easy to foresee the drone industry splitting into maker and service provider similar to the robotics industry of today. In robotics, a few big firms make the basic industrial robot arms but thousands of independent global consultants, distributors, integrators and engineers add value through software and add-on devices. In the drone world the service provider segment is growing as fast as the maker group. Mapping, data mining, autopilot providers (with or separate from collision avoidance systems), data analysis, sensor manufacturers and aerial service companies are beginning to do today for the drone industry what integrators, distributors and consulting engineers have done for the robotics industry in the past.

https://youtu.be/mh9cYgWdKIY

https://youtu.be/q8o_R0WAwGE