Robocars

NHTSA ODI report exonerates Tesla in fatal crash

NHTSA released the report from their Office of Defects Investigation on the fatal Tesla crash in Florida last spring. It’s a report that is surprisingly favorable to Tesla. So much so that even I am surprised. While I did not think Tesla would be found defective, this report seems to come from a different agency than the one that recently warned comma.ai that:

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.

The ODI report rules that Tesla properly considered driver distraction risks in its design of the product. It goes even further, noting that drivers using Tesla autopilot (including those monitoring it properly and those who did not) still had a decently lower accident rate for mile than drivers of ordinary cars without autopilot. In other words, while the autopilot without supervision is not good enough to drive on its own, the autopilot even with occasionally lapsed supervision that is known to happen is still overall a safer system than not having the autopilot at all.

This will provide powerful support for companies developing autopilot style systems, and companies designing robocars who wish to use customer supervised driving as a means to build up test miles and verification data. They are not putting their customers at risk as long as they do it as well as Tesla. This is interesting (and the report notes that evaluation of autopilot distraction is not a settled question) because it seems probable that people using the autopilot and ignoring the road to do e-Mail or watch movies are not safer than regular drivers. But the overall collection of distracted and watchful drivers is still a win.

This might change as companies introduce technologies which watch drivers and keep them out of the more dangerous inattentive style of use. As the autopilots get better, it will become more and more tempting, after all.

Tesla stock did not seem to be moved by this report. But it was also not moved by the accident or other investigations — it actually went on a broadly upward course for 2 months following announcement of the fatality.

The ODI’s job is to judge if a vehicle is defective. That is different from saying it’s not perfect. Perfection is not expected, especially from ADAS and similar systems. The discussion about the finer points of whether drivers might over-trust the system are not firmly settled here. That can still be true without the car being defective and failing to perform as designed, or being designed negligently.

Rodney Brooks on Pedestrian Interaction, Andrew Ng on Infrastructure, and both on human attitudes

Recently we’ve seen two essays by people I highly respect in the field of AI and robotics. Their points are worthy of reading, but in spite of my respect, I have some differences of course.

The first essay comes from Andrew Ng, head of AI (and thus the self-driving car project) at Baidu. You will find few who can compete with Andrew when it comes to expertise on AI. (Update: This essay is not recent, but I only came upon it recently.)

In Wired he writes that Self-Driving Cars Won’t Work Until We Change Our Roads—And Attitudes. And the media have read this essay as being much more strong about changing the roads than he actually writes. I have declared it to be the “first law of robocars” that you don’t change the infrastructure. You improve your car to match the world you are given, you don’t ask the world to change to help your cars. There are several reasons I promote this rule:

  • As soon as you depend on a change in the world in order to drive safely, you have vastly limited where you can deploy. You declare that your technology will be, for a very long time, a limited area technology.
  • You have to depend on, and wait for others to change the world or their attitudes. It’s beyond your control.
  • When it comes to cities and infrastructure, the pace of change is glacial. When it comes to human behaviour, it can be even worse.
  • While it may seem that the change to infrastructure is clearer and easier to plan, the reality is almost assuredly the opposite. That’s because the clever teams of developers, armed with the constantly improving technologies driven by Moore’s law, have the ability to solve problems in a way that is much faster than our linear intuitions suggest. Consider measuring traffic by installing tons of sensors, vs. just getting everybody to download Waze. Before Waze, the sensor approach seemed clear, if expensive. But it was wrong.

As noted, Andrew Ng does not actually suggest that much change to the infrastructure. He talks about:

  • Having road construction crews log changes to the road before they do them
  • Giving police and others who direct traffic a more reliable way to communicate their commands to cars
  • Better painting of lane markers
  • More reliable ways to learn the state of traffic lights
  • Tools to help humans understand the actions and plans of robocars

Logging construction

The first proposal is one I have also made, because it’s very doable, thanks to computer technology. All it requires at first blush is a smartphone app in the hands of construction crews. Before starting a project, they would know that just as important as laying out cones and signs is opening the app and declaring the start of a project. The phone has a GPS and can offer a selection of precise road locations and log it. Of course, the projects should be logged even before they begin, but because that’s imperfect, smartphone logging is good enough. You could improve this by sticking old smartphones in all the road construction machines (old phones are cheap and there are only so many machines) so that any time a machine stops on a road for very long, it sends a message to a control center. Even emergency construction gets detected this way.

Even with all that, cars still need to detect changes to the road (that’s easy with good maps) and cones and machines. Which they can do.

Police redirection

I think the redirection problem is more difficult. Many people redirect traffic, even civilians. However, I would be interested to see Ng’s prediction on how hard it is to get neural network based recognizers to understand all the common gestures. Considering that computers are now getting better at reading sign languages, which are much more complex, I am optimistic here. But in any event, there is another solution for the cases where the system can’t understand the advice, namely calling in an operator in a remote control center, which is what Nissan plans to do, and what we do at Starship. Unmanned cars, with no human to help, will just avoid data dead zones. If somehow they get to them, there can be other solutions, which are imperfect but fine when the problem is very rare, such as a way for the traffic manager to speak to the car (after all, spoken language understanding is now close to a solved problem for limited vocabulary problems.)

Better lanes

Here I disagree with Andrew. His statement may be a result of efforts to drive on roads without maps, even though Baidu has good map expertise. Google’s car has a map of the texture of the road. It knows where the cracks and jagged lane markers are. The car actually likes degrading lane markers. It’s perfectly painted straight and smooth roads which confuse it (though only slightly, and not enough to cause a problem.) So no, I think that better line painting is not on the must-do list.

Seeing lights

He’s right, seeing lights can be challenging, though the better cars are getting good at it. The simple algorithm is “you don’t go if you don’t confirm green.” That means you don’t run a red but you could block traffic. If that’s very rare it’s OK. We can consider infrastructure to solve that, though I’m wary. Fortunately, if the city is controlling its lights with a central computer, you don’t have to alter the traffic light itself (which is hard,) you can just query the city, in those rare cases, for when the light will be changing. I think that problem will be solved, but I also think it may well be solved just by better cameras. Good robocars know exactly where all the lights are, and they know where they are, and thus they know exactly what pixels in a video image are from the light, even if the sun is behind it. (Good robocars also know where the sun is and will avoid stopping in a place where there is no light they can see without the sun right behind it.)

Working with people

How cars interact with people is one of Andrew Ng’s points and the central point of Rodney Brooks’ essay Unexpected Consequences of Self Driving Cars. Already many of the car companies have had fun experimenting with that, putting displays on the outside of cars of various sorts. While cars don’t have the body language and eye contact of human drivers, I don’t predict a problem we can’t solve with good effort.

Brooks’ credentials are also superb, as founder of iRobot (Roomba) and Rethink Robotics (Baxter) as well as many accomplishments as an MIT professor. His essay delves into one of the key questions I have wondered about for some time — how to deal with a world where things do not follow the rules, and where there are lots of implicit and changing rules and interactions. Google discovered the first instant of this when their car got stuck at a 4 way stop by being polite. They had to program the car to assert its right to go in order to handle the stop. Likewise, you need to speed to be a good citizen on many of our roads today.

His key points are as follows:

  • There is a well worked out dance between pedestrians and cars, that varies greatly among different road types, with give and take, and it’s not suitable for machines yet.
  • People want to know a driver has seen them before stepping near or certainly in front of a vehicle.
  • People jaywalk, and even expect cars to stop for them when they do on some streets.
  • In snowy places, people walk on the street when the sidewalk is not shoveled.
  • Foot traffic can be so much that timid cars can’t ever get out of sidestreets or driveways. Nice pedestrians often let them out. They will hand signal their willingness to yield or use body language.
  • Sometimes people just stand at the corner or edge of the road, and you can’t tell if they are standing there or getting ready to cross.
  • People setting cars to circle rather than park
  • People might jump out of their car to do something, leaving it in the middle of the street blocking traffic, where today they would be unwilling to double park.
  • People might abuse parking spots by having a car “hold” them for quick service when they want to leave an event.
  • Cars will grab early spots to pick up children at schools.

Brooks starts with one common mistake — he has bought into the “levels” defined by SAE, even claiming them to be well accepted. In fact, many people don’t accept them, especially the most advanced developers, and I outlined recently why there is only one level, namely unmanned operation, and so the levels are useless as a taxonomy. Instead the real taxonomy in the early days will be the difference between mobility on demand services (robotaxi) and self-drive enabled high end luxury cars. Many of his problems involve privately owned cars and selfish behaviour by their owners. Many of those behaviours don’t make sense in a world with robotaxis. I think it’s very likely that the robotaxis come first, and come in large numbers first, while some imagine it’s the other way around.

Unintended consequences

Brooks is right that there will be unintended consequences, and the technology will be put to uses nobody thought of. People will be greedy, and antisocial, that can be assured. Fortunately, however, people will work out solutions, in advance, to anything you can think of or notice just by walking down the street or thinking about issues for a few days. The experienced developers have been thinking about these problems for decades now, and cars like Google’s have driven for 300 human lifetimes of driving, and that number keeps increasing. They note every unusual situation they encounter on every road they can try to drive, and the put it into the simulator if it’s important. They’ve already seen more situations than any one human will encounter on those roads, though they certainly haven’t driven all the types of road in the world. But they will, before they certify as safe for deployment on such roads.

As I noted, only the “level 4” situation is real. Level 5 is an aspirational science-fiction goal, and the others are unsafe. Key to the improved thinking on “levels” it is no longer the amount of human supervision needed that makes the difference, it is the types of roads and situations you can handle. All these vehicles will only handle a subset of roads, and that is what everybody plans. If there is a road that is too hard, they just won’t drive it. Fortunately, there are lots of road subsets out there that are very, very useful and make economic sense. For a while, many companies planned only to do highways, which are the simplest road subset of all, except for the speed. A small subset, but everybody agrees it’s valuable.

So the short answer is, solutions will be found to these problems if the roads they occur on are commercially necessary. If they are not necessary, the solutions will be delayed until they can be found, though that’s probably not too long.

As noted above, many people do expect systems to be developed to allow dialogue between robocars and pedestrians or other humans. One useful tool is gaze detection — just as a cheap flash camera causes “red eye” in photos, machines shining infrared light can easily tell if you are looking at them. Eye contact in that direction is detectable. There have been various experiments in sending information in the reverse direction. Some cars have lasers that can paint lines on the road. Others can display text. Some have an LED ribbon surrounding them that shows all the objects and people tracked by the car, so people can understand that they are being perceived. You can also flash a light back directly at people to return their eye contact — I see you and I see that you saw me.

Over time, we’ll develop styles of communication, and they will get standarized. It’s not essential to do that on day one; you just stay on the simpler roads until you know you can handle the others. Private cars will pause and pop out a steering wheel. Services like Uber will send you a human driver in the early days if the car is going somewhere the systems can’t drive, or they might even let you drive part of it. Such incrementalism is the only way it can ever work.

People taking advantage of timidity of robocars

I believe there are solutions to some of the problems laid out. One I have considered is pedestrians and others who take advantage of the naturally conservative and timid nature of a robocar. If people feel they can safely cut off or jaywalk in front of robocars, they will. And the unmanned cars will mostly just accept that, though only about 10% of all cars should be unmanned at any given time. The cars with passengers are another story. Those passengers will be bothered if they are cut off, or forced to brake quickly. They will spill their coffee. And they will fight back.

Citizen based strong traffic code enforcement

Every time you jump in front of such a car, it will of course have saved the video and other sensor data. It’s always doing that. But the passenger might tell the car, “Please save that recent encounter. E-mail it to the police.” The police will do little with it at first, but in time, especially since there are rich people in these cars, they will throw a face recognizer and licence plate recognizer on the system that gets the videos. They will notice that one person keeps jaywalking right in front of the cars and annoying the passengers. Or the guy who keeps cutting off the cars as though they are not there because they always brake. They will have video of him doing it 40 times, or 100. And at that point, they will do something. The worst offender will get identified and get an E-mail from police. We have 50 videos of you doing this. Here are 50 tickets. Then the next, and the next until nobody wants to get to the top of the list.

This might actually create pressure the other way — a street that belongs only to the cars and excludes the non-car user. A traffic code that is enforced to the letter because every person inconvenienced has an ability to file a complaint trivially. We don’t want that either, but we can control that balance.

I actually look forward to fixing one of the dynamics of jaywalking that doesn’t work. Often, people like to jaywalk and a car is approaching. They want to have the car pass at full speed and then walk behind it — everybody is more comfortable behind a car than in front of one. But the driver gets paranoid and stops, and eventually you uncomfortably cross in front, annoyed at that and that you stopped somebody you didn’t intend to stop. I suspect robocars will be able to handle this dynamic better, predicting when people might actually be on a path to enter their lane, but not slowing down for stopped pedestrians (adults at least) and trust them to manage their crossing. Children are a different matter.

People being selfish with robocars

Brooks wonders about people doing selfish things with their robocars. Here, he mostly talks about privately owned robocars, since most of what he describes would not or could not happen with a robotaxi. There will be some private cars so we want to think about this.

A very common supposition I see here and elsewhere is the idea of a car that circles rather than parking. Today, operating a car is about $20/hour so that’s already completely irrational, and even when robocar operation drops to $8/hour or less, parking is going to be ridiculously cheap and plentiful so that’s not too likely. There could be competition for spots in very busy areas (schools, arenas etc.) which don’t have much space for pick-up and drop-off, and that’s another area where a bit of traffic code could go a long way. Allow facilities to make a rule: “No car may enter unless its passenger is waiting at the pick-up spot” with authority to ticket and evict any car that does otherwise. Over time, such locations will adjust their pick-up spots to the robocar world and become more like Singapore’s airport, which provides amazing taxi throughput with no cab lines by making it all happen in parallel. Of course, cars would wait outside the zone but robocars can easily double and triple park without blocking the cars they sit in the path of. Robocars waiting for passengers at busy locations will be able to purchase waiting spaces for less than the cost of circling, and then serve their customers or owners. If necessary, market prices can be put on the prized close waiting spaces to solve any problems of scarcity.

So when can it happen?

Robocars will come to different places at different times. They will handle different classes of streets at different times. They will handle different types of interactions with pedestrians and other road users at different times. Where you live will dictate when you can use it and how you can use it. Vendors will push at the most lucrative routes to start, then work down. There will be many problems that are difficult at first, and the result will be the early cars just don’t go on those sorts of streets or into those sorts of situations. Human driving, either by the customer or something like an Uber driver, will fill in the gaps.

Long before then, teams will have encountered or thought of just about any situation you’ve seen, and any situation you’ve likely thought of in a short amount of time. They will have programmed every variation of that situation they can imagine into their simulators to see what their car does. They will use this to grow the network of roads the cars handle every day. Even if at the start, it is not a network of use to you, it won’t be too long before it becomes that, at first for some of your rides, and eventually for most or all.

CES News Part 1 -- cars

CES has become the big event for major car makers to show off robocar technology. Most of the north hall, and a giant and valuable parking lot next to it, were devoted to car technology and self-driving demos.

Gallery of CES comments

Earlier I posted about many of the pre-CES announcements and it turns out there were not too many extra events during the show. I went to visit many of the booths and demos and prepared some photo galleries. The first is my gallery on cars. In this gallery, each picture has a caption so you need to page through them to see the actual commentary at the bottom under the photo. Just 3 of many of the photos are in this post.

To the left you see BMW’s concept car, which starts to express the idea of an ultimate non-driving machine. Inside you see that the back seat has a bookshelf in it. Chances are you will just use your eReader, but this expresses and important message — that the car of the future will be more like a living, playing or working space than a transportation space.

Nissan

The main announcement during the show was from Nissan, which outlined their plans and revealed some concept cars you will see in the gallery. The primary demo they showed involved integration of some technology worked on by Nissan’s silicon valley lab leader, Maarten Sierhuis in his prior role at NASA. Nissan is located close to NASA Ames (I myself work at Singularity University on the NASA grounds) and did testing there.

Their demo showed an ability to ask a remote control center to assist a car with a situation it doesn’t understand. When the car sees something it can’t handle, it stops or pulls over, and people in the remote call center can draw a path on their console to tell the car where to go instead. For example, it can be drawn how to get around an obstacle, or take a detour, or obey somebody directing traffic. If the same problem happens again, and it is approved, the next car can use the same path if it remains clear.

I have seen this technology a number of places before, including of course the Mars rovers, and we use something like it at Starship Technologies for our delivery robots. This is the first deployment by a major automaker.

Nissan also committed to deployment in early 2020 as they have before — but now it’s closer.

You can also see Nissan’s more unusual concepts, with tiny sensor pods instead of side-view mirrors, and steering wheels that fold up.

Startups

Several startups were present. One is AIMotive, from Hungary. They gave me a demo ride in their test car. They are building a complete software suite, primarily using cameras and radar but also able to use LIDAR. They are working to sell it to automotive OEMs and already work with Volvo on DriveMe. The system uses neural networks for perception, but more traditional coding for path planning and other functions. It wasn’t too fond of Las Vegas roads, because the lane markers are not painted there — lanes are divided only with Bott’s Dots. But it was still able to drive by finding the edge of the road. They claim they now have 120 engineers working on self-driving systems in Hungary.  read more »

No, a Tesla didn't predict an accident and brake for it

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 articles most assuredly were not saying the Tesla predicted the accident that never happened had the Tesla failed to brake, they are talking about predicting the dramatic crash shown in the video.)

The accident is brutal but apparently nobody was hurt.

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.  read more »

Ford Fusion, Nvidia, MobilEye, HERE, Intel and other partnerships, BMW 7s in 2017 and more CES pre-news

Thursday night I am heading off to CES, and it’s become the main show it seems for announcing robocar news. There’s already a bunch.

BMW says it will deploy a fleet of 40 cars in late 2017

Bumping up the timetables, BMW has declared it will have a fleet of 40 self-driving series 7 cars, using BMW’s technology combined with MobilEye and Intel. Intel has recently been making a push to catch up to Nvidia as a chipmaker supplier to automakers for self-driving. It’s not quite said what the cars will do, but they will be trying lots of different roads. So far BMW has mostly been developing its own tech. More interesting has been their announcement of plans to sell rides via their DriveNow service. This was spoken of a year ago but not much more has been said.

Intel also bought 15% of “HERE” the company formerly known as Navteq and Nokia. Last year, the German automakers banded together to buy HERE from Nokia and the focus has been on “HD” self-driving maps.

Hyundai, Delphi show off cars

There are demo cars out there from Delphi and a Hyundai Ioniq. Delphi’s car has been working for a while (it’s an Audi SUV) but recently they have also added a bunch of MobilEye sensors to it. Reports about the car are good, and they hope to have it ready by 2019, showing up in 2020 or 2021 cars on dealer lots.

Toyota sticks to concepts

Toyota’s main announcement is the Concept-i meant to show off some UI design ideas. It’s cute but still very much a car, though with all the silly hallmarks of a concept — hidden wheels, strangely opening doors and more.

Quanergy announces manufacturing plans for $250 solid state LIDAR

Quanergy (Note: I am on their advisory board) announced it will begin manufacturing this year of automotive grade $250 solid state LIDARs. Perhaps this will stop all the constant articles about how LIDAR is super-expensive and means that robocars must be super-expensive too. The first model is only a taste of what’s to come in the next couple of years as well.

New Ford Model has sleeker design

Ford has become the US carmaker to watch (in addition to Tesla) with their announcement last year that they don’t plan to sell their robocars, only use them to offer ride service in fleets. They are the first and only carmaker to say this is their exclusive plan. Just prior to CES, Ford showed off a new test model featuring smaller Velodyne pucks and a more deliberate design.

I have personally never understood the desire to design robocars to “look like regular cars.” I strongly believe that, just like the Prius, riders in the early robocars will want them to look distinctive, so they can show off how they are in a car of the future. Ford’s carm based on the Fusion hybrid, is a nice compromise — clearly a robocar with its sensors, but also one of sleek and deliberate design.

Nvidia keeps its push

Nvidia has a new test car they have called BB8. (Do they have to licence that name?) It looks fairly basic, and they show a demo of it taking somebody for a ride with voice control, handling a lot of environments. It’s notable that at the end, the driver has to take over to get to the destination, so it doesn’t have everything, nor would we expect it. NVIDIA is pushing their multi-GPU board as the answer to how to get a lot of computing power to run neural networks in the car.

What’s coming

Announcements are due tomorrow from Nissan and probably others. I’ll report Friday from the show floor. See you there.

DMV pulls out big guns, Uber backs off, Waymo minivans emerge & Honda next?

The California DMV got serious in their battle with Uber and revoked the car registrations for Uber’s test vehicles. Uber had declined to register the cars for autonomous testing, using an exemption in that law which I described earlier. The DMV decided to go the next step and pull the more basic licence plate every car has to have if based in California. Uber announced it would take the cars to another state.

While I’m friends with the Uber team, I have not discussed this matter with them, so I can only speculate why it came to this. As noted, Uber was complying with the letter of the law but not the spirit, which the DMV didn’t like. At the same time, the DMV kept pointing out that registering was really not that hard or expensive, so they can’t figure out why Uber stuck to its guns. (Of course, Uber has a long history of doing that when it comes to cities trying to impose old-world taxi regulations on them.)

The DMV is right, it’s not hard to register. But with that registration comes other burdens, in particular filing regular public reports on distance traveled, interventions and any accidents. Companies doing breakthrough R&D don’t usually work under such regimes, and I am speculating this might have been one of Uber’s big issues. We’ve all see the tremendous amount of press that Google has gotten over accidents which were clearly not the fault of their system. The question is whether the public’s right to know (or the government’s) about risks to public safety supersedes the developer’s desires to keep their research projects proprietary and secret.

It’s clear that we would not want a developer going out on the roads and having above-average numbers of accidents and keeping it hidden. And it may also be true that we can’t trust the developers to judge the cause of fault, because they could have a bias. (Though on most of the teams I have seen, the bias has been a safety paranoid one, not the other way around.)

Certainly when we let teens start to drive, we don’t have them make a public report of any accidents they have. The police and DMV know, and people who get too many tickets or accidents get demerits and lose licences when it is clear they are a danger to the public. Perhaps a reasonable compromise would have been that all developers report all problems to the DMV, but that those results are not made public immediately. They would be revealed eventually, and immediately if it was determined the system was at fault.

Uber must be somewhat jealous of Tesla. Tesla registered several cars under the DMV system, and last I saw, they sent in their reports saying their cars had driven zero miles. That’s because they are making use of the same exemption that Uber wanted to make use of, and saying that the cars are not currently qualifying as autonomous under the law.

Pacifica Minivans released by Waymo/Google

Waymo has unveiled its 4th generation vehicle, a fleet of 100 Chrysler Pacifica minivans. The partnership was announced in May and the new vehicle has various custom modifications

As you can see, the van still has Waymo’s custom 360 degree LIDAR dome on top, and two sensors at the back top corners, plus other forward sensors. The back sensors I would guess to be rear radar — which lets you make lane changes safely. We also see three apparent small LIDARs, one on the front bumper, and the other two on the sides near the windshield pillars with what may be side-view radars.

A bumper LIDAR makes sure you can see what’s right in front of the bumper, an area that the rooftop LIDAR might not see. That’s important for low speed operations and parking, or situations where there might be something surprising right up close. I am reminded of reports from the Navya team that when they deployed their shuttles, teens would try to lie down in front of the shuttle to find out if it would stop for them. Teens will be teens, so you may need a sensor for that.

Side radar is important for cross traffic when trying to do things like making turns at stop signs onto streets with high speed. Google also has longer range LIDAR to help with that.

The minivan is of course the opposite end of the spectrum from the 2-passenger no-steering-wheel 3rd generation prototype. That car tested many ideas for low speed urban taxi operations, and the new vehicle seems aimed at highway travel and group travel (with six or more seats.) One thing people particularly like is that like most minivans these days, it has an automatic sliding door. Somehow that conveys the idea of a robotic taxi even more when it opens the door for you! The step-in-step-out convenience of the minivan does indeed give people a better understanding of the world of frictionless transportation that is coming.

Honda Next?

Update: Also announced yesterday was a partnership between Honda and Waymo. It says they will be putting the Waymo self-driving system into Honda cars. While the details in the release are scant, this actually could be a much bigger announcement than the minivans, in which Chrysler’s participation is quite minimal. Waymo has put out the spec for the modified minivan, and Chrysler builds it to their spec, then Waymo installs the tech. A Waymo vehicle sourced from Chrysler. The Honda release suggests something much bigger — a Honda vehicle sourced from, or partnering with Waymo.

There has not been as much press about this Honda announcement but it may be the biggest one.

NPRM for DSRC and V2V

The DoT has finally released their proposed rules requiring all new cars (starting between 2020 and 2022) to come equipped with vehicle-to-vehicle radio units, speaking the DSRC protocol and blabbing their location everywhere they go. Regular readers will know that I think this is a pretty silly idea, even a dangerous one from the standpoint of privacy and security, and that most developers of self-driving cars, rather than saying this is a vital step, describe it as “something we would use if it gets out there, but certainly not essential for our vehicles.”

I will have more comment on this in the future.

Uber's battle in San Francisco

For a few months, Uber has been testing their self-driving prototypes in Pittsburgh, giving rides to willing customers with a safety driver (or two) in the front seat monitoring the drive and ready to take over.

When Uber came to do this in San Francisco, starting this week, it was a good step to study new territory and new customers, but the real wrinkle was they decided not to get autonomous vehicle test permits from the California DMV. Google/Waymo and most others have such permits. Telsa has such permits but claims it never uses them.

I played an advisory role for Google when the Nevada law was drafted, and this followed into the California law. One of the provisions in both laws is that they specifically exempt cars that are unable to drive without a human supervisor. This provision showed up, not because of the efforts of Google or other self-drive teams, but because the big automakers wanted to make sure that these new self-driving laws did not constrain the only things they were making at the time — advanced ADAS and “autopilot” cars which are effectively extra-fancy cruise controls that combine lanekeeping functions with adaptive cruise control for speed. Many car makers offered products like that going back a decade, and they wanted to make sure that whatever crazy companies like Google wanted in their self-driving laws, it would not pertain to them.

The law says:

“…excluding vehicles equipped with one or more systems that enhance safety or provide driver assistance but are not capable of driving or operating the vehicle without the active physical control or monitoring of a natural person.”

Now Uber (whose team is managed by my friend Anthony Levandowski who played a role in the creation of those state laws while he was at Google) wants to make use of these carve-outs to do their pilot project. As long as their car is tweaked so that it can’t drive without human monitoring, it would seem to fit under that exemption. (I don’t know, but would presume they might do some minor modifications so the system can’t drive without the driver weight sensor activated, or a button held down or similar to prove the driver is monitoring.)

The DMV looks at it another way. Since their testing regulations say you can’t test without human safety drivers monitoring and ready to take over, it was never the intent of the law to effectively exempt everything. You can’t test a car without human monitoring under the regulations, but cars that need monitoring are exempt. The key is calling the system a driver assistance system rather than a driving system.

The DMV is right about the spirit. Uber may be right about the letter. Of course, Uber has a long history of not being all that diligent in complying with the law, and then getting the law to improve, but this time, I think they are within the letter. At least for a while.

Other News

Velodyne reports success in research into solid state LIDAR. Velodyne has owned the market for self-driving car LIDAR for years, as they are the only producers of a high-end model. Their models are mechanical and very expensive, so other companies have been pushing the lower cost end of the market, including Quanergy (Where I am an advisor) which has also had solid state LIDAR for some time, and appears closer to production.

These and others verify something that most in the industry have expected for some time — LIDAR is going to get cheap soon. Companies like Tesla, which have avoided LIDAR because you can’t get a decently priced unit in production quantities, have effectively bet that cameras will get good before LIDAR gets cheap. The reality is that most early cars will simply use both cheap LIDAR and improving neural network based vision at the same time.

Google car is now Waymo

Google’s car project (known as “Chauffeur”) really kickstarted the entire robocar revolution, and Google has put in more work, for longer, than anybody. The car was also the first project of what became Google “X” (or just “X” today under Alphabet. Inside X, a lab devoted to big audacious “moonshot” projects that affect the physical world as well as the digital, they have promoted the idea that projects should eventually “graduate,” moving from being research to real commercial efforts.

Alphabet has announced that the project will be its own subsidiary company with the new name “Waymo.” The name is not the news, though; what’s important is the move away from being a unit of a mega-company like Google or Alphabet. The freedoms to act that come with being a start-up (though a fairly large and well funded one) are greater than units in large corporations have. Contrast what Uber was able to do, skirting and even violating the law until it got the law changed, with what big corporations need to do.

Google also released information about how in 2015 they took Steve Mahan — the blind man who was also the first non-employee to try out a car for running errands — for the first non-employee (blind or otherwise) fully self-driving ride on public streets, in a vehicle with no steering wheel and no backup safety driver in the vehicle. (This may be an effort to counter the large amount of press about public ride offerings by Nutonomy in Singapore and Uber in Pittsburgh, as well as truck deliveries by Uber/Otto in 2016.)

It took Google/Alphabet 6 years to let somebody ride on public streets in part because it is a big company. It’s an interesting contrast with how Otto did a demonstration video after just a few months of life of a truck driving a Nevada highway with nobody behind the wheel (but Otto employees inside and around it.) That’s the sort of radical step that startups.

Waymo has declared their next goal is to “let people use our vehicles to do everyday things like run errands, commute to work, or get safely home after a night on the town.” This is the brass ring, a “Mobility on Demand” service able to pick people up (ie. run unmanned) and even carry a drunk person.

The last point is important. To carry a drunk is a particular challenge. In terms of improving road safety it’s one of the most worthwhile things we could do with self-driving cars, since drunks have so many of the accidents. To carry a drunk, you can’t let the human take control even if they want to. Unlike unmanned operation, you must travel at the speed impatient humans demand, and you must protect the precious cargo. To make things worse, in some legal jurisdictions, they still want to consider the person inside the car the “driver,” which could mean that since the “driver” is impaired, operation is illegal.

Waymo as leader

The importance of this project is hard to overstate. While most car companies had small backburner projects related to self-driving going back many years, and a number of worthwhile research milestones were conquered in the 90s and even earlier, the Google/Waymo project, which sprang from the Darpa Grand Challenge, energized everybody. Tiny projects at car companies all got internal funding because car companies couldn’t tolerate the press and the world thinking and writing the that true future of the car was coming from a non-car company, a search engine company. Now the car companies have divisions with thousands of engineers, and it’s thanks to Google. The Google/Waymo team was accomplishing tasks 5 years ago that most projects are only now just getting to, especially in non-highway driving. They were rejecting avenues (like driving with a human on standby ready to take the wheel on short notice) in 2013 that many projects are still trying to figure out.

Indeed, even in 2010, when I first joined the project and it had just over a dozen people, it had already accomplished more complex tasks that most projects, even the Tesla autopilot that some people think is in the lead, have yet to accomplish.

Let’s see where Waymo goes.

What disability rules are right for robotaxis?

Robocars are broadly going to be a huge boon for many people with disabilities, especially disabilities which make it difficult to drive or those that make it hard to get in and out of vehicles. Existing disability regulations and policies were written without robocars in mind, and there are probably some improvements that need to be made.

While I was at Google, I helped slightly with the project to show the first non-employee getting to use the car to run errands. The subject we selected was 95% blind, and of course he can’t drive, and even using transit is a burden. It was obvious to him immediately how life-changing the technology will be.

Some background on disabled transport

There are two rough policy approaches to making things more accessible. One requires that we make everything accessible. The other uses special accommodations for the disabled.

Making everything accessible is broadly preferred by advocates. Wheelchair ramps on all public buildngs etc. Doing less than this runs a risk of “separate but equal” which quickly becomes separate and inferior. It’s also hugely expensive, and while that cost is borne by people like building owners and society, there is not unlimited budget, and there are arguments that there may be more efficient ways to spend the resources that are available. There are also lots of very different disabilities, and you need very different methods to deal with impairments in sight, mobility, hearing, cognition and the rest.

Over 50 million people in the USA have some sort of disability, so this is no minor matter.

In transportation, there is a general goal to make public transit accessible. To supplement that, or where that is not done, there are the paratransit rules. Paratransit offers people who meet certain tests an alternate ride (usually in a door to door van) for themselves and a helper for no more than twice the cost of a regular bus ticket. That sounds great until you learn you also have to schedule it a day in advance, and have a one-hour pickup window (which the disabled hate) and it’s hugely expensive, with an average cost per ride of over $30, which cities hate. (In the worst towns, it is $60/ride.) In some cities it approaches half the transit budget. Some cities, looking at that huge cost, let some disabled customers just use taxis for short trips, which provide much better service and cost much less. (Though to avoid over-use they put limitations on this.)

There are Americans with Disabilities Act rules for taxis. Regular sedan taxis are not directly regulated though there can be no discrimination of disabled customers who are capable of riding in a sedan. Any new van of up to 8 seats has to been accessible, which often means things like wheelchair lifts. In addition, once a taxi fleet has accessible vans, it has to offer “equivalent service” levels. This might mean that if it has 200 sedans, it can’t buy just one van because there would be much longer wait times to get that van. To get around this, a lot of companies use a loophole and purchase only used vans. The law only covers the use of new vans. Companies like Uber and Lyft don’t own vehicles at all, and so are not governed in the same way by fleet requirements, though they do offer accessible vehicle services in some cities.

When Uber and similar companies move to offering robotaxi service with vehicles they own, these laws would apply to them. Unlike some companies, the used van loophole will also be difficult since most robotaxis will be custom built new.

New Types of Vehicles

Robotaxi service offers the promise of a vehicle on demand, and it offers the potential of a vehicle well fitted to the trip. Mostly I talk about things like the ability to use a small and inexpensive one person vehicle for solo urban trips (which are 80% of trips, so this is a big deal) but it also means sending an SUV when 3 people want to go skiing, or a pickup-truck for a work run, or a van designed for socializing when a group of people want to travel together.

It also offers the ability to create vehicles just for people with certain disabilities. One example I find quite interesting is the Kenguru — a small, single person vehicle which is hollow, and allows a user in a wheelchair to just roll in the back and steer it with hand controls. For wheelchair users with working arms, this is hugely superior to designs that require you to get out of your chair into a car seat, or which involve the time delays of using a wheelchair lift. Especially with nobody to assist. Roll-in, roll-out can match the convenience of the able-bodied. The current Kenguru is to be steered, but a self-driving vehicle like this could handle even those in power chairs, and offer a fold-down bench for an able-bodied companion.

Being computerized, these vehicles will also offer accessible user interfaces. Indeed, they may mostly rely on the user’s phone, which will already be customized to their needs.

Custom-designed to meet particular disabilities, these vehicles will both serve the disabled better and frankly be not that useful for others. As such, regimes that require adapting all vehicles to handle both types of customers may have the right spirit, but provide inferior service.

Another key benefit of robotaxi service for the disabled will be the low price. Reduced job prospects drive many with disabilities into poverty. Service that is naturally low in price will be enabling.

Equivalent service or Separate but Superior

Providing “equivalent” service is difficult with traditional taxis, particularly for smaller fleets. Robotaxis, which don’t mind waiting around because no human driver is waiting, make this easier to do. The service level of a robotaxi service is based on the density of currently unused vehicles in your area. Increase fleet size with the same demand, and service level goes up. As long as fleet size is not way overblown, so that vehicles still wear out by the mile rather than by the year, increasing fleet size is not nearly as expensive as it is for regular cars or human driven taxis.

This means you can, fairly readily, offer equivalent or even superior service at a pretty reasonable cost. As long as disabled-designed vehicles are made in decent quantities to keep their costs low, the cost should be close to the cost of regular vehicles. In the public interest, regular vehicle customers might subsidize the slightly higher cost of these lower volume vehicles.

With increased fleets, service levels would generally be superior to the regular fleets, but not always. The law generally allows this, but the disabled community will need to understand a few unequal things that probably will happen:

  • Slightly more advanced notice of rides will often make it possible to provide service at lower cost. Regular vehicles will naturally be present on every block. Disabled vehicles might be present with less density during high use times, but the ability to reposition lets even slight advance notice do a lot.
  • For those in groups, it may not be easy to carry a person in a wheelchair along with several non-wheelchair passengers. This might mean the wheelchair passenger goes in their own vehicle (with videoconference link.) This is not as good, but is much more cost effective than requiring every van to have a wheelchair lift.
  • To increase service levels, it is likely competing companies would cooperate on serving the disabled, and pool fleets. Until the disabled become a profitable market rather than one done to meet goals of public good, companies will prefer to work together. As such if you call for an Uber, you might often get a Lyft or other small fleet car.
  • Low cost disabled transport may mean that accessible public transit and paratransit slowly fade. Public transit which has its own tracks will continue to be accessible as it offers a speed advantage which may not be met on the roads, but otherwise it may be much cheaper to offer private robotaxis than to make all transit accessible. This would mean a group of people might not be able to ride transit together if it’s not accessible.
  • Small electric vehicles may be allowed to enter buildings, dropping passengers right at elevator lobbies or other destinations.

The biggest trade-off will be the loss of social group experiences. There certainly will be buses and vans with lifts which allow groups of mixed-ability passengers to travel together, but it is unlikely these would be so common as to offer the same service level as ordinary vans. With advance notice of just 10 minutes, they could probably be available.

What if the city ran Waze and you had to obey it? Could this cure congestion?

I believe we have the potential to eliminate a major fraction of traffic congestion in the near future, using technology that exists today which will be cheap in the future. The method has been outlined by myself and others in the past, but here I offer an alternate way to explain it which may help crystallize it in people’s minds.

Today many people drive almost all the time guided by their smartphone, using navigation apps like Google Maps, Apple Maps or Waze (now owned by Google.) Many have come to drive as though they were a robot under the command of the app, trusting and obeying it at every turn. Tools like these apps are even causing controversy, because in the hunt for the quickest trip, they are often finding creative routes that bypass congested major roads for local streets that used to be lightly used.

Put simply, the answer to traffic congestion might be, “What if you, by law, had to obey your navigation app at rush hour?” To be more specific, what if the cities and towns that own the streets handed out reservations for routes on those streets to you via those apps, and your navigation app directed you down them? And what if the cities made sure there were never more cars put on a piece of road than it had capacity to handle? (The city would not literally run Waze, it would hand out route reservations to it, and it would still do the UI and be a private company.)

The value is huge. Estimates suggest congestion costs around 160 billion dollars per year in the USA, including 3 billion gallons of fuel and 42 hours of time for every driver. Roughly quadruple that for the world.

Road metering actually works

This approach would exploit one principle in road management that’s been most effective in reducing congestion, namely road metering. The majority of traffic congestion is caused, no surprise, by excess traffic — more cars trying to use a stretch of road than it has the capacity to handle. There are other things that cause congestion — accidents, gridlock and irrational driver behaviour, but even these only cause traffic jams when the road is near or over capacity.

Today, in many cities, highway metering is keeping the highways flowing far better than they used to. When highways stall, the metering lights stop cars from entering the freeway as fast as they want. You get frustrated waiting at the metering light but the reward is you eventually get on a freeway that’s not as badly overloaded.

Another type of metering is called congestion pricing. Pioneered in Singapore, these systems place a toll on driving in the most congested areas, typically the downtown cores at rush hour. They are also used in London, Milan, Stockholm and some smaller towns, but have never caught on in many other areas for political reasons. Congestion charging can easily be viewed as allocating the roads to the rich when they were paid for by everybody’s taxes.

A third successful metering system is the High-occupancy toll lane. HOT lanes take carpool lanes that are being underutilized, and let drivers pay a market-based price to use them solo. The price is set to bring in just enough solo drivers to avoid wasting the spare capacity of the lane without overloading it. Taking those solo drivers out of the other lanes improves their flow as well. While not every city will admit it, carpool lanes themselves have not been a success. 90% of the carpools in them are families or others who would have carpooled anyway. The 10% “induced” carpools are great, but if the carpool lane only runs at 50% capacity, it ends up causing more congestion than it saves. HOT is a metering system that fixes that problem.  read more »

Comma One goes Open Source, Robocars in New Zealand Earthquakes and more

There have been few postings this month since I took the time to enjoy a holiday in New Zealand around speaking at the SingularityU New Zealand summit in Christchurch. The night before the summit, we enjoyed a 7.8 earthquake not so far from Christchurch, whose downtown was over 2/3 demolished after quakes in 2010 and 2011. On the 11th floor of the hotel, it was a disturbing nailbiter of swaying back and forth for over 2 minutes — but of course swaying is what the building is supposed to do; that means it’s working. The shocks were rolling, not violent, and in fact we got more violent jolts from aftershocks a week later when we went to Picton.

While driving around that region, we encountered this classic earthquake scene on the road:

There were many like this, and in fact the main highway of the South Island was destroyed long-term not too far away, cutting off several towns. A scene like this makes you wonder just what a robocar would do in such situations. I already answered this question in a blog post on how to handle a tsunami. Fortunately there was only a mild tsunami for this quake. A tsunami will result in a warning in the rich world, and the car will know the elevation map of the roads and know how to get to high ground. In some places, like Japan,t here is also an advanced earthquake warning system that tells you quakes are coming well before they hit you, since electrons go much faster than seismic waves. With such a system, robocars should receive a warning and come to a stop unless they need to evacuate a tsunami zone. Without such a warning, we still could imagine the road cracking and collapsing in front of you as might have happened on this road. Of course the cones and signs that warned me days later would not be present.

The answer again lies in the fact that pictures like mine will be used to create situations like this in simulator, and all car developers will be able to test their systems with simulated quake damage to make sure they do the right thing. I’ve spoken since 2010 on the value of a shared simulator environment and I think if government agencies like NHTSA want to really help development, providing funding and tools for such an environment would be a good step. NHTSA’s proposal that all developers share their logs of all incidents would clearly make such a simulator better, but there is pushback because of the proprietary value of those logs. When it comes to strange situations like earthquakes, I doubt there would be much pushback on having an open and shared simulator environment.

New Zealand’s government is taking a very welcoming approach to robocars. They are not regulating for a while, and have invited developers to come and test. They have even said it’s OK to test unmanned vehicles under some fairly simple rules. NZ does not have any auto industry, and of course it’s quite remote, but we’ll see if they can attract developers to come test. Their roads feature something you don’t see much in the USA — tons and tons of one-lane bridges and other one-lane stretches of highway. Turns out that robocars, with a little bit of communication, can make very superhumanly efficient use of one-lane two-way roads, and it might be worth exploring.

Open Source Comma One box

Speaking of Open, today Comma.ai, which previously had declared they were giving up on their neural network autopilot due to NHTSA threats today announced they have open sourced their software, along with hardware designs and case designs. NHTSA did not want them making an autopilot, and said they could not simply rely on the fact that drivers were told they must be diligent, it will be very interesting to see how NHTSA reacts to the release of open designs that anybody can then install on their car.

The automotive industry has had a long history of valuing the tinkerer. All the big car companies had their beginnings with small tinkerers and inventors. Some even died in the very machines they were inventing. These beginnings have allowed people to do all sorts of playing around in their garages with new car ideas, without government oversight, in spite of the risk to themselves and even others on the road. If a mechanic wants to charge you for working on your car, they must be licenced, but you are free to work on it yourself with no licence, and even build experimental cars. You just can’t sell them. And even those rights have been eroded.

Clearly far fewer people will have the inclination to build an autopilot using the comma.ai tools by themselves. But it won’t be that hard to do, and they can make it easier with time, too. One could even imagine a car which already had the necessary hardware, so that you only needed to download software to make it happen.

In recent times, there has been a strong effort to prevent people with tinkering with their cars, even in software. One common area of controversy has been around engine tuning. Engine tuning is regulated by the EPA to keep emissions low. Car vendors have to show they have done this — and they can’t program their car to give good emissions only on the test while getting better performance off the test as VW did. But owners have been known to want to make such modifications. Now we will see modifications that affect not just emissions but safety. Car companies don’t want to be responsible if you modify the code in your car and there is an accident involving both their code and yours. As such, they will try to secure their car systems so you can’t change them, and the government may help them or even insist on it. When you add computer security risks to the mix — who can certify the modified car can’t be taken over and used as a weapon? — it will get even more fun.

I will also point out that I suspect that comma’s approach would not know what to do about the collapsed road, because it would never have been trained in that situation. It might, however, simply sound an alert and kick out, not being able to find the lane any more.

Regulatory pushback

Regular readers will have seen my strong critique of the NHTSA rules. The other major news during my break was the pushback from major players in the public comment on the regulations. In some ways the regulations didn’t do enough to give vendors the certainty they need to make their plans. At the same time, they were criticsed for not giving enough flexibility to vendors. In addition, as expected, they resist giving up their proprietary data in the proposed forced sharing. I predict continued ambivalence on the regulations. Big players actually like having lots of regulations, because big players know how to deal with that and small players don’t.

Comma.ai cancels comma-one add-on box after threats from NHTSA

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 which would also apply to Tesla and other automakers which have built supervised autopilot functions.
  2. Of particular interest is the paragraph which 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 it seems that 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.

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 yeterday 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 element of the regulations, which should not be forbidding this approach until we know it to be dangerous.  read more »

Of the SAE's robocar "levels" only level 4 will be meaningful, and only partly

It’s no secret that I’ve been a critic of the NHTSA “levels” as a taxonomy for types of Robocars since the start. Recent changes in their use calls for some new analysis that concludes that only one of the levels is actually interesting, and only tells part of the story at that. As such, they have become even less useful as a taxonomy. Levels 2 and 3 are unsafe, and Level 5 is remote future technology. Level 4 is the only interesting one and there is thus no taxonomy.

Unfortunately, they have just been encoded into law, which is very much the wrong direction.

NHTSA and SAE both created a similar set of levels, and they were so similar that NHTSA declared they would just defer to the SAE’s system. Nothing wrong with that, but the core flaws are not addressed by this. Far better, their regulations declared that the levels were just part of the story, and they put extra emphasis on what they called the “operating domain” — namely what locations, road types and road conditions the vehicle operates in.

The levels focus entirely on the question of how much human supervision a vehicle needs. This is an important issue, but the levels treated it like the only issue, and it may not even be the most important. My other main criticism was that the levels, by being numbered, imply a progression for the technology. That progression is far from certain and in fact almost certainly wrong. SAE updated its levels to say that they are not intended to imply a progression, but as long as they are numbers this is how people read them.

Today I will go further. All but level 4 are uninteresting. Some may never exist, or exist only temporarily. They will be at best footnotes of history, not core elements of a taxonomy.

Level 4 is what I would call a vehicle capable of “unmanned” operation — driving with nobody inside. This enables most of the interesting applications of robocars.

Here’s why the other levels are less interesting:

Levels 0 and 1 — Manual or ADAS-improved

Levels 0 and 1 refer to existing technology. We don’t really need new terms for our old cars. Level 2 perhaps best described as a more advanced version of level 1 and that transition has already taken place.

Level 2 — Supervised Autopilot

Supervised autopilots are real. This is what Tesla sells, and many others have similar offerings. They are working in one of two ways. The first is the intended way, with full time supervision. This is little more than a more advanced cruise control, and may not even be as relaxing.

The second way is what we’ve seen happen with Tesla — a car that needs supervision, but is so good at driving that supervisors get complacent and stop supervising. They want a full self-driving car but don’t have it, so they pretend they do. Many are now saying that this makes the idea of supervised autopilot too dangerous to deploy. The better you make it, the more likely it can lull people into bad activity.

Update: One day after I wrote this, it was revealed that NHTSA shut down comma.ai’s efforts to build an aftermarket autopilot citing these concerns, among others.

Level 3 — Standby driver

This level is really a variation of Level 4, but the vehicle needs the ability to call upon a driver who is not paying attention and get them to take control with 10 to 60 seconds of advance warning. Many people don’t think this can be done safely. When Google experimented with it in 2013, they concluded it was not safe, and decided to take the steering wheel entirely out of their experimental vehicles.

Even if Level 3 is a real thing, it will be short lived as people seek an unmanned capable vehicle. And Level 4 vehicles will offer controls for special use, even if they don’t permit a transition while moving.

Level 5 — Drive absolutely everywhere

SAE, unlike NHTSA’s first proposal, did want to make it clear that an unmanned capable (Level 4) vehicle would only operate in certain places or situations. So they added level 5 to make it clear that level 4 was limited in domain. That’s good, but the reality is that a vehicle that can truly drive everywhere is not on anybody’s plan. It probably requires AI that matches human beings.

Consider this situation in which I’ve been driven. In the African bush on a game safari, we spot a leopard crossing the road. So the guide drives the car off-road (on private land) running over young trees, over rocks, down into wet and dry streambeds to follow the leopard. Great fun, but this is unlikely to be an ability there is ever market demand to develop. Likewise, there are lots of small off-road tracks that are used by only one person. There is no economic incentive for a company to solve this problem any time soon.

Someday we might see cars that can do these things under the high-level control a human, but they are not going to do them on their own, unmanned. As such SAE level 5 is academic, and serves only to remind us that level 4 does not mean everywhere.

Levels vs. Cul-de-sacs

The levels are not a progression. I will contend in fact that even to the extent that levels 2, 3/4 and 5 exist, they are quite probably entirely different technologies.

Level 2 is being done with ADAS technologies. They are designed to have a driver in the loop. Their designs in many case do not have a path to the reliability level needed for unmanned, which is orders of magnitude higher. It is not just a difference of degree, it is one of kind.

Level 3 is related to level 4, in particular because a level 3 car is expected to be able to handle non-response from its driver, and safely stop or pull off the road. It can be viewed as a sucky version of a level 4 system. (It’s also not that different — see below.)

Level 5, as indicated, probably requires technologies that are more like artificial general intelligence than they are like a driving system.

As such the levels are not levels. There is no path between any of the levels and the one above it, except in the case of 3/4.

Level 4

This leaves Level 4 as the only one worth working on long term, the only one with talking about. The others are just there to create a contrast. NHTSA realizes this and gave the name ODD (Operational Design Domain) to refer to the real area of research, namely what roads and situations the vehicles can handle.

The distinction between 4 and 3 is also not as big as you might expect. Google removed the steering wheel from their prototype to set a high bar for themselves, but they actually left one in for use in testing and development. In reality, even the future’s unmanned cars will feature some way in which a human can control them, for use during breakdowns, special situations, and moving the cars outside of their service areas (operational domains.) Even if the transition from autodrive to human drive is unsafe at speed, it will still be safe if the car pulls over and activates the controls for a licenced driver.

As such, the only distinction of a “level 3” car is it hopes to be able to do that transition while moving, on short but not urgent notice. A pretty minor distinction to be a core element of a taxonomy.

If Level 4 is the only interesting one, my recommendation is to drop the levels from our taxonomy, and focus the taxonomy instead on the classes of roads and conditions the vehicle can handle. It can be a given that outside of those operating domains, other forms of operation might be used, but that does not bear much on the actual problem.

I say we just identify a vehicle capable of unmanned or unsupervised operation as a self-driving car or robocar, and then get to work on the real taxonomy of problems.

Yikes - even Barack Obama wants to solve robocar "Trolley Problems" now

I had hoped I was done ranting about our obsession with what robocars will do in no-win “who do I hit?” situations, but this week, even Barack Obama in his interview with Wired opined on the issue, prompted by my friend Joi Ito from the MIT Media Lab. (The Media Lab recently ran a misleading exercise asking people to pretend they were a self-driving car deciding who to run over.)

I’ve written about the trouble with these problems and even proposed a solution but it seems there is still lots of need to revisit this. Let’s examine why this problem is definitely not important enough to merit the attention of the President or his regulators, and how it might even make the world more dangerous.

We are completely fascinated by this problem

Almost never do I give a robocar talk without somebody asking about this. Two nights ago, I attended another speaker’s talk and he got the question as his 2nd one. He looked at his watch and declared he had won a bet with himself about how quickly somebody would ask. It has become the #1 question in the mind of the public, and even Presidents.

It is not hard to understand why. Life or death issues are morbidly attractive to us, and the issue of machines making life or death decisions is doubly fascinating. It’s been the subject of academic debates and fiction for decades, and now it appears to be a real question. For those who love these sorts of issues, and even those who don’t, the pull is inescapable.

At the same time, even the biggest fan of these questions, stepping back a bit, would agree they are of only modest importance. They might not agree with the very low priority that I assign, but I don’t think anybody feels they are anywhere close to the #1 question out there. As such we must realize we are very poor at judging the importance of these problems. So each person who has not already done so needs to look at how much importance they assign, and put an automatic discount on this. This is hard to do. We are really terrible at statistics sometimes, and dealing with probabilities of risk. We worry much more about the risks of a terrorist attack on a plane flight than we do about the drive to the airport, but that’s entirely wrong. This is one of those situations, and while people are free to judge risks incorrectly, academics and regulators must not.

Academics call this the Law of triviality. A real world example is terrorism. The risk of that is very small, but we make immense efforts to prevent it and far smaller efforts to fight much larger risks.

These situations are quite rare, and we need data about how rare they are

In order to judge the importance of these risks, it would be great if we had real data. All traffic fatalities are documented in fairly good detail, as are many accidents. A worthwhile academic project would be to figure out just how frequent these incidents are. I suspect they are extremely infrequent, especially ones involving fatality. Right now fatalities happen about every 2 million hours of driving, and the majority of those are single car fatalities (with fatigue and alcohol among leading causes.) I have still yet to read a report of a fatality or serious injury that involved a driver having no escape, but the ability to choose what they hit with different choices leading to injuries for different people. I am not saying they don’t exist, but first examinations suggest they are quite rare. Probably hundreds of billions of miles, if not more, between them.

Those who want to claim they are important have the duty to show that they are more common than these intuitions suggest. Frankly, I think if there were accidents where the driver made a deliberate decision to run down one person to save another, or to hurt themselves to save another, this would be a fairly big human interest news story. Our fascination with this question demands it. Just how many lives would be really saved if cars made the “right” decision about who to hit in the tiny handful of accidents where they must hit somebody?

In addition, there are two broad classes of situations. In one, the accident is the fault of another party or cause, and in the other, it is the fault of the driver making the “who to hit” decision. In the former case, the law puts no blame on you for who you hit if forced into the situation by another driver. In the latter case, we have the unusual situation that a car is somehow out of control or making a major mistake and yet still has the ability to steer to hit the “right” target.

These situations will be much rarer for robocars

Unlike humans, robocars will drive conservatively and be designed to avoid failures. For example, in the MIT study, the scenario was often a car whose brakes had failed. That won’t happen to robocars — ever. I really mean never. Robocar designs now all commonly feature two redundant braking systems, because they can’t rely on a human pumping the hydraulics manually or pulling an emergency brake. In addition, every time they apply the brakes, they will be testing them, and at the first sign of any problem they will go in for repair. The same is true of the two redundant steering systems. Complete failure should be ridiculously unlikely.

The cars will not suddenly come upon a crosswalk full of people with no time to stop — they know where the crosswalks are and they won’t drive so fast as to not be able to stop for one. They will be also constantly measuring traction and road conditions to assure they don’t drive too fast for the road. They won’t go around blind corners at high speeds. They will have maps showing all known bottlenecks and construction zones. Ideally new construction zones will only get created after a worker has logged the zone on their mobile phone and the updates are pushed out to cars going that way, but if for some reason the workers don’t do that, the first car to encounter the anomaly will make sure all other cars know.

This does not mean the cars will be perfect, but they won’t be hitting people because they were reckless or had predictable mechanical failures. Their failures will be more strange, and also make it less likely the vehicle will have the ability to choose who to hit.

To be fair, robocars also introduce one other big difference. Humans can argue that they don’t have time to think through what they might do in a split-second accident decision. That’s why when they do hit things, we call them accidents. They clearly didn’t intend the result. Robocars do have the time to think about it, and their programmers, if demanded to by the law, have the time to think about it. Trolley problems demand the car be programmed to hit something deliberately. The impact will not be an accident, even if the cause was. This puts a much higher standard on the actions of the robocar. One could even argue it’s an unfair standard, which will delay deployment if we need to wait for it.

In spite of what people describe in scenarios, these cars won’t leave their right of way

It is often imagined an ethical robocar might veer into the oncoming lane or onto the sidewalk to hit a lesser target instead of a more vulnerable one in its path. That’s not impossible, but it’s pretty unlikely. For one, that’s super-duper illegal. I don’t see a company, unless forced to do so, programming a car to ever deliberately leave its right of way in order to hit somebody. It doesn’t matter if you save 3 school buses full of kids, deliberately killing anybody standing on the sidewalk sounds like a company-ruining move.

For one thing, developers just won’t put that much energy into making their car drive well on the sidewalk or in oncoming traffic. They should not put their energies there! This means the cars will not be well tested or designed when doing this. Humans are general thinkers, we can handle driving on the grass even though we have had little practice. Robots don’t quite work that way, even ones designed with machine learning.

This limits most of the situations to ones where you have a choice of targets within your right-of-way. And changing lanes is always more risky than staying in your lane, especially if there is something else in the lane you want to change to. Swerving if the other lane is clear makes sense, but swerving into an occupied lane is once again something that is going to be uncharted territory for the car.

By and large the law already has an answer

The vehicle code is quite detailed about who has right-of-way. In almost every accident, somebody didn’t have it and is the one at fault under the law. The first instinct for most programmers will be to have their car follow the law and stick to their ROW. To deliberately leave your ROW is a very risky move as outlined above. You might get criticized for running over jaywalkers when you could have veered onto the sidewalk, but the former won’t be punished by the law and the latter can be. If people don’t like the law, they should change the law.

The lesson of the Trolley problem is “you probably should not try to solve trolley problems.”

Ethicists point out correctly that Trolley problems may be academic exercises, but are worth investigating for what they teach. That’s true in the classroom. But look at what they teach! From a pure “save the most people” utilitarian standpoint, the answer is easy — switch the car onto the track to kill one in order to save 5. But most people don’t pick that answer, particularly in the “big man” version where you can push a big man standing with you on a bridge onto the tracks to stop the trolley and save the 5. The problem teaches us we feel much better about leaving things as they are than in overtly deciding to kill a bystander. What the academic exercise teaches us is that in the real world, we should not foist this problem on the developers.

If it’s rare and a no-win situation, do you have to solve it?

Trolley problems are philosophy class exercises to help academics discuss ethical and moral problems. They aren’t guides to real life. In the classic “trolley problem” we forget that none of it happens unless a truly evil person has tied people to a railway track. In reality, many would argue that the actors in a trolley problem are absolved of moral responsibility because the true blame is on the setting and its architect, not them. In philosophy class, we can still debate which situation is more or less moral, but they are all evil. These are “no win” situations, and in fact one of the purposes of the problems is they often describe situations where there is no clear right answer. All answers are wrong, and people disagree about which is most wrong.

If a situation is rare, and it takes effort to figure out which is the less wrong answer, and things will still be wrong after you do this even if you do it well, does it make sense to demand an answer at all? To individuals involved, yes, but not to society. The hard truth is that with 1.2 million auto fatalities a year — a number we all want to see go down greatly — it doesn’t matter that much to society whether, in a scenario that happens once every few years, you kill 2 people or 3 while arguing which choice was more moral. That’s because answering the question, and implementing the answer, have a cost.

Every life matters, but we regularly make decisions like this. We find things that are bad and rare, and we decide that below a certain risk threshold, we will not try to solve them unless the cost is truly zero. And here the cost is very far from zero. Because these are no-win situations and each choice is wrong, each choice comes with risk. You may work hard to pick the “right” choice and end up having others declare it wrong — all to make a very tiny improvement in safety.

At a minimum each solution will involve thought and programming, as well as emotional strain for those involved. It will involve legal review and in the new regulations, certification processes and documentation. All things that go into the decision must be recorded and justified. All of this is untrod legal ground making it even harder. In addition, no real scenario with match hypothetical situations exactly, so the software must apply to a range of situations and still do the intended thing (let alone the right thing) as the situation varies. This is not minor.

Nobody wants to solve it

In spite of the fascination these problems hold, coming up with “solutions” to these no-win situations are the last things developers want to do. In articles about these problems, we almost always see the statement, “Who should decide who the car will hit?” The answer is nobody wants to decide. The answer is almost surely wrong in the view of some. Nobody is going to get much satisfaction or any kudos for doing a good job, whatever that is. Combined with the rarity of these events compared to the many other problems on the table, solving ethical issues is very, very, very low on the priority list for most teams. Because developers and vendors don’t want to solve these questions and take the blame for those solutions, it makes more sense to ask policymakers to solve what needs to be solved. As Christophe von Hugo of Mercedes put it, “99% of our engineering work is to prevent these situations from happening at all.”

The cost of solving may be much higher than people estimate

People grossly underestimate how hard some of these problems will be to solve. Many of the situations I have seen proposed actually demand that cars develop entirely new capabilities that they don’t need except to solve these problems. In these cases, we are talking about serious cost, and delays to deployment if it is judged necessary to solve these problems. Since robocars are planned as a life-saving technology, each day of delay has serious consequences. Real people will be hurt because of these delays aimed at making a better decision in rare hypothetical situations.

Let’s consider some of the things I have seen:

  • Many situations involve counting the occupants of other cars, or counting pedestrians. Robocars don’t otherwise have to do this, nor can they easily do it. Today it doesn’t matter if there are 2 or 3 pedestrians — the only rule is not to hit any number of pedestrians. With low resolution LIDAR or radar, such counts are very difficult. Counts inside vehicles are even harder.
  • One scenario considers evaluating motorcyclists based on whether they are wearing helmets. I think this one is ridiculous, but if people take it seriously it is indeed serious. This is almost impossible to discern from a LIDAR image and can be challenging even with computer vision.
  • Some scenarios involve driving off cliffs or onto sidewalks or otherwise off the road. Most cars make heavy use of maps to drive, but they have no reason to make maps of off-road areas at the level of detail that goes into the roads.
  • More extreme scenarios compare things like children vs. adults, or school-buses vs. regular ones. Today’s robocars have no reason to tell these apart. And how do you tell a dwarf adult from a child? Full handling of these moral valuations requires human level perception in some cases.
  • Some suggestions have asked cars to compare levels of injury. Cars might be asked to judge the difference between a fatal impact and one that just breaks a leg.

These are just a few examples. A large fraction of the hypothetical situations I have seen demand some capability of the cars that they don’t have or don’t need to have just to drive safely.

The problem of course is there are those who say that one must not put cars on the road until the ethical dilemmas have been addressed. Not everybody says this but it’s a very common sentiment, and now the new regulations demand at least some evaluation of it. No matter how much the regulations might claim they are voluntary, this is a false claim, and not just because some states are already talking about making them more mandatory.

Once a duty of care has been suggested, especially by the government, you ignore it at your peril. Once you know the government — all the way to the President — wants you to solve something, then you must be afraid you will be asked “why didn’t you solve that one?” You have to come up with an answer to that, even with voluntary compliance.

The math on this is worth understanding. Robocars will be deployed slowly into society but that doesn’t matter for this calculation. If robocars are rare, they can prevent only a smaller number of accidents, but they will also encounter a correspondingly smaller number of trolley problems. What matters is how many trolley situations there are per fatality, and how many people you can save with better handling of those problems. If you get one trolley problem for every 1,000 or 10,000 fatalities, and robocars are having half the fatalities, the math very clearly says you should not accept any delay to work on these problems.

The court of public opinion

The real courts may or may not punish vendors for picking the wrong solution (or the default solution of staying in your lane) in no-win situations. Chances are there will be a greater fear of the court of public opinion. There is reason to fear the public would not react well if a vehicle could have made an obviously better outcome, particularly if the bad outcome involves children or highly vulnerable road users vs. adults and at-fault or protected road users.

Because of this I think that many companies will still try to solve some of these problems even if the law puts no duty on them. Those companies can evaluate the risk on their own and decide how best to mitigate it. That should be their decision.

For a long time, many people felt any robocar fatality would cause uproar in the public eye. To everybody’s surprise, the first Tesla autopilot deaths resulted in Tesla stock rising for 2 months, even with 3 different agencies doing investigations. While the reality of the Tesla is that the drivers bear much more responsibility than a full robocar would, the public isn’t very clear on that point, so the lack of reaction is astonishing. I suspect companies will discount this risk somewhat after this event.

This is a version 2 feature, not a version 1 feature

As noted, while humans make split-second “gut” decisions and we call the results accidents, robocars are much more intentional. If we demand they solve these problems, we ask something of them and their programmers that we don’t ask of human drivers. We want robocars to drive more safely than humans, but we also must accept that the first robocars to be deployed will only be a little better. The goal is to start saving lives and to get better and better at it as time goes by. We must consider the ethics of making the problem even harder on day one. Robocars will be superhuman in many ways, but primarily at doing the things humans do, only better. In the future, we should demand these cars meet an even higher standard than we put on people. But not today: The dawn of this technology is the wrong time to also demand entirely new capabilities for rare situations.

Performing to the best moral standards in rare situations is not something that belongs on the feature list for the first cars. Solving trolley situations well is in the “how do we make this perfect?” problem set, not the “how do we make this great?” set. It is important to remember how the perfect can be the enemy of the good and to distinguish between the two. Yes, it means accepting there are low chance that somebody could be hurt or die, but people are already being killed, in large numbers, by the human drivers we aim to replace.

So let’s solve trolley problems, but do it after we get the cars out on the road both saving lives and teaching us how to improve them further.

What about the fascination?

The over-fascination with this problem is a real thing even if the problem isn’t. Studies have displayed one interesting result after surveying people: When you ask people what a car should do for the good of society, they would want it to sacrifice its passenger to save multiple pedestrians, especially children. On the other hand if you ask people if they would buy a car that did that, far fewer said yes. As long as the problem is rare, there is no actual “good of society” priority; the real “good of society” comes from getting this technology deployed and driving safely as quickly as possible. Mercedes recently announced a much simpler strategy which does what people actually want, and got criticism for it. Their strategy is reasonable — they want to save the party they can be most sure of saving, namely the passengers. They note that they have very little reliable information on what will happen in other cars or who is in them, so they should focus not on a guess of what would save the most people, but what will surely save the people they know about.

What should we do?

I make the following concrete recommendations:

  1. We should do research to determine how frequent these problems are, how many have “obvious” answers and thus learn just how many fatalities and injuries might be prevented by better handling of these situations.
  2. We should remove all expectation on first generation vehicles that they put any effort into solving the rare ones, which may well be all of them.
  3. It should be made clear there is no duty of care to go to extraordinary lengths (including building new perception capabilities) to deal with sufficiently rare problems.
  4. Due to the public over-fascination, vendors may decide to declare their approaches to satisfy the public. Simple approaches should be encouraged, at in the early years of this technology, almost no answer should be “wrong.”
  5. For non-rare problems, governments should set up a system where developers/vendors can ask for rulings on the right behaviour from the policymakers, and limit the duty of care to following those rulings.
  6. As the technology matures, and new perception abilities come online, more discussion of these questions can be warranted. This belongs in car 2.0, not car 1.0.
  7. More focus at all levels should go into the real everyday ethical issues of robocars, such as roads where getting around requires regularly violating the law (speeding, aggression etc.) in the way all human users already do.
  8. People writing about these problems should emphasize how rare they are, and when doing artificial scenarios, recount how artificial they are. Because of the public’s fears and poor risk analysis, it is inappropriate to feed on those fears rather than be realistic.

NHTSA Regulations part 4: Crashes, Training, Certification, State Law, Operation, Validation and Autopilots

After my initial reactions and Overall Analysis here is a point by point consideration of second set of elements from NHTSA’s 15 point certification list for robocars. See my series for other articles or the first half of the list.

Crashworthiness

In this section, the remind vendors they still need to meet the same standards as regular cars do. We are not ready to start removing heavy passive safety systems just because the vehicles get in fewer crashes. In the future we might want to change that, as those systems can be 1/3 of the weight of a vehicle.

They also note that different seating configurations (like rear facing seats) need to protect as well. It’s already the case that rear facing seats will likely be better in forward collisions. Face-to-face seating may present some challenges in this environment, as it is less clear how to deploy the airbags. Taxis in London often feature face-to-face seating, though that is less common in the USA. Will this be possible under these regulations?

The rules also call for unmanned vehicles to absorb energy like existing vehicles. I don’t know if this is a requirement on unusual vehicle design for regular cars or not. (If it were, it would have prohibited SUVs with their high bodies that can cause a bad impact with a low-body sports-car.)

Consumer Education and Training

This seems like another mild goal, but we don’t want a world where you can’t ride in a taxi unless you are certified as having taking a training course. Especially if it’s one for which you have very little to do. These rules are written more for people buying a car (for whom training can make sense) than those just planning to be a passenger.

Registration and Certification

This section imagines labels for drivers. It’s pretty silly and not very practical. Is a car going to have a sticker saying “This car can drive itself on Elm St. south of Pine, or on highway 101 except in Gilroy?” There should be another way, not labels, that this is communicated, especially because it will change all the time.

Post-Crash Behavior

This set is fairly reasonable — it requires a process describing what you do to a vehicle after a crash before it goes back into service.

Federal, State and Local Laws

This section calls for a detailed plan on how to assure compliance with all the laws. Interestingly, it also asks for a plan on how the vehicle will violate laws that human drivers sometimes violate. This is one of the areas where regulatory effort is necessary, because strictly cars are not allowed to violate the law — doing things like crossing the double-yellow line to pass a car blocking your path.  read more »

NHTSA Regulations part 3: Data Sharing, Privacy, Safety, Security and HMI

After my initial reactions and Overall Analysis here is a point by point consideration of the elements from NHTSA’s 15 point certification list for robocars. See also the second half and the whole series

Let’s dig in:

Data Recording and Sharing

These regulations require a plan about how the vehicle keep logs around any incident (while following privacy rules.) This is something everybody already does — in fact they keep logs of everything for now — since they want to debug any problems they encounter. NHTSA wants the logs to be available to NHTSA for crash investigation.

NHTSA also wants recordings of positive events (the system avoided a problem.)

Most interesting is a requirement for a data sharing plan. NHTSA wants companies to share their logs with their competitors in the event of incidents and important non-incidents, like near misses or detection of difficult objects.

This is perhaps the most interesting element of the plan, but it has seen some resistance from vendors. And it is indeed something that might not happen at scale without regulation. Many teams will consider their set of test data to be part of their crown jewels. Such test data is only gathered by spending many millions of dollars to send drivers out on the roads, or by convincing customers or others to voluntarily supervise while their cars gather test data, as Tesla has done. A large part of the head-start that leaders have in this field is the amount of different road situations they have been able to expose their vehicles to. Recordings of mundane driving activity are less exciting and will be easier to gather. Real world incidents are rare and gold for testing. The sharing is not as golden, because each vehicle will have different sensors, located in different places, so it will not be easy to adapt logs from one vehicle directly to another. While a vehicle system can play its own raw logs back directly to see how it performs in the same situation, other vehicles won’t readily do that.

Instead this offers the ability to build something that all vendors want and need, and the world needs, which is a high quality simulator where cars can be tested against real world recordings and entirely synthetic events. The data sharing requirement will allow the input of all these situations into the simulator, so every car can test how it would have performed. This simulation will mostly be at the “post perception level” where the car has (roughly) identified all the things on the road and is figuring out what to do with them, but some simulation could be done at lower levels.

These data logs and simulator scenarios will create what is known as a regression test suite. You test your car in all the situations, and every time you modify the software, you test that your modifications didn’t break something that used to work. It’s an essential tool.

In the history of software, there have been shared public test suites (often sourced from academia) and private ones that are closely guarded. For some time, I have proposed that it might be very useful if there were a a public and open source simulator environment which all teams could contribute scenarios to, but I always expected most contributions would come from academics and the open source community. Without this rule, the teams with the most test miles under their belts might be less willing to contribute.

Such a simulator would help all teams and level the playing field. It would allow small innovators to even build and test prototype ideas entirely in simulator, with very low cost and zero risk compared to building it in physical hardware.

This is a great example of where NHTSA could use its money rather than its regulatory power to improve safety, by funding the development of such test tools. In fact, if done open source, the agencies and academic institutions of the world could fund a global one. (This would face opposition from companies hoping to sell test tools, but there will still be openings for proprietary test tools.)

Privacy

This section demands a privacy policy. I’m not against that, though of course the history of privacy policies is not a great one. They mostly involve people clicking “I agree” to things they don’t read. More important is the requirement that vendors be thinking about privacy.

The requirement for user choice is an interesting one, and it conflicts with the logging requirements. People are wary of technology that will betray them in court. Of course, as long as the car is not a hybrid car that mixes human driving with self-driving, and the passenger is not liable in an accident, there should be minimal risk to the passenger from accidents being recorded.

The rules require that personal information be scrubbed from any published data. This is a good idea but history shows it is remarkably hard to do properly.  read more »

Detailed analysis of NHTSA robocar regulations: Overview

The recent Federal Automated Vehicles Policy is long. (My same-day analysis is here and the whole series is being released.) At 116 pages (to be fair, less than half is policy declarations and the rest is plans for the future and associated materials) it is much larger than many of us were expecting.

The policy was introduced with a letter attributed to President Obama, where he wrote:

There are always those who argue that government should stay out of free enterprise entirely, but I think most Americans would agree we still need rules to keep our air and water clean, and our food and medicine safe. That’s the general principle here. What’s more, the quickest way to slam the brakes on innovation is for the public to lose confidence in the safety of new technologies. Both government and industry have a responsibility to make sure that doesn’t happen. And make no mistake: If a self-driving car isn’t safe, we have the authority to pull it off the road. We won’t hesitate to protect the American public’s safety.

This leads in to an unprecedented effort to write regulations for a technology that barely exists and has not been deployed beyond the testing stage. The history of automotive regulation has been the opposite, and so this is a major change. The key question is what justifies such a big change, and the cost that will come with it.

Make no mistake, the cost will be real. The cost of regulations is rarely known in advance but it is rarely small. Regulations slow all players down and make them more cautious — indeed it is sometimes their goal to cause that caution. Regulations result in projects needing “compliance departments” and the establishment of procedures and legal teams to assure they are complied with. In almost all cases, regulations punish small companies and startups more than they punish big players. In some cases, big players even welcome regulation, both because it slows down competitors and innovators, and because they usually also have skilled governmental affairs teams and lobbying teams which are able to subtly bend the regulations to match their needs.

This need not even be nefarious, though it often is. Companies that can devote a large team to dealing with regulations, those who can always send staff to meetings and negotiations and public comment sessions will naturally do better than those which can’t.

The US has had a history of regulating after the fact. Of being the place where “if it’s not been forbidden, it’s permitted.” This is what has allowed many of the most advanced robocar projects to flourish in the USA.

The attitude has been that industry (and startups) should lead and innovate. Only if the companies start doing something wrong or harmful, and market forces won’t stop them from being that way, is it time for the regulators to step in and make the errant companies do better. This approach has worked far better than the idea that regulators would attempt to understand a product or technology before it is deployed, imagine how it might go wrong, and make rules to keep the companies in line before any of them have shown evidence of crossing a line.

In spite of all I have written here, the robocar industry is still young. There are startups yet to be born which will develop new ideas yet to be imagined that change how everybody thinks about robocars and transportation. These innovative teams will develop new concepts of what it means to be safe and how to make things safe. Their ideas will be obvious only well after the fact.

Regulations and standards don’t deal well with that. They can only encode conventional wisdom. “Best practices” are really “the best we knew before the innovators came.” Innovators don’t ignore the old wisdom willy-nilly, they often ignore it or supersede it quite deliberately.

What’s good?

Some players — notably the big ones — have lauded these regulations. Big players, like car companies, Google, Uber and others have a reason to prefer regulations over a wild west landscape. Big companies like certainty. They need to know that if they build a product, that it will be legal to sell it. They can handle the cost of complex regulations, as long as they know they can build it.  read more »

Critique of NHTSA's newly released regulations

The long awaited list of recommendations and potential regulations for Robocars has just been released by NHTSA, the federal agency that regulates car safety and safety issues in car manufacture. Normally, NHTSA does not regulate car technology before it is released into the market, and the agency, while it says it is wary of slowing down this safety-increasing technology, has decided to do the unprecedented — and at a whopping 115 pages.

Broadly, this is very much the wrong direction. Nobody — not Google, Uber, Ford, GM or certainly NHTSA — knows the precise form of these cars will have when deployed. Almost surely something will change from our existing knowledge today. They know this, but still wish to move. Some of the larger players have pushed for regulation. Big companies like certainty. They want to know what the rules will be before they invest. Startups thrive better in the chaos, making up the rules as we go along.

NHTSA hopes to define “best practices” but the best anybody can do in 2016 is lay down existing practices and conventional wisdom. The entirely new methods of providing safety that are yet to be invented won’t be in such a definition.

The document is very detailed, so it will generate several blog posts of analysis. Here I present just initial reactions. Those reactions are broadly negative. This document is too detailed by an order of magnitude. Its regulations begin today, but fortunately they are also accepting public comment. The scope of the document is so large, however, that it seems extremely unlikely that they would scale back this document to the level it should be at. As such, the progress of robocar development in the USA may be seriously negatively affected.

Vehicle performance guidelines

The first part of the regulations is a proposed 15 point safety standard. It must be certified (by the vendor) that the car meets these standards. NHTSA wants the power, according to an Op-Ed by no less than President Obama, to be able to pull cars from the road that don’t meet these safety promises.

  • Data Recording and Sharing
  • Privacy
  • System Safety
  • Vehicle Cybersecurity
  • Human Machine Interface
  • Crashworthiness
  • Consumer Education and Training
  • Registration and Certification
  • Post-Crash Behavior
  • Federal, State and Local Laws
  • Operational Design Domain
  • Object and Event Detection and Response
  • Fall Back (Minimal Risk Condition)
  • Validation Methods
  • Ethical Considerations

As you might guess, the most disturbing is the last one. As I have written many times, the issue of ethical “trolley problems” where cars must decide between killing one person or another are a philosophy class tool, not a guide to real world situations. Developers should spend as close to zero effort on these problems as possible, since they are not common enough to warrant special attention, if not for our morbid fascination with machines making life or death decisions in hypothetical situations. Let the policymakers answer these questions if they want to; programmers and vendors don’t.

For the past couple of years, this has been a game that’s kept people entertained and ethicists employed. The idea that government regulations might demand solutions to these problems before these cars can go on the road is appalling. If these regulations are written this way, we will delay saving lots of real lives in the interest of debating which highly hypothetical lives will be saved or harmed in ridiculously rare situations.

NHTSA’s rules demand that ethical decisions be “made consciously and intentionally.” Algorithms must be “transparent” and based on input from regulators, drivers, passengers and road users. While the section makes mention of machine learning techniques, it seems in the same breath to forbid them.

Most of the other rules are more innocuous. Of course all vendors will know and have little trouble listing what roads their car works on, and they will have extensive testing data on the car’s perception system and how it handles every sort of failure. However, the requirement to keep the government constantly updated will be burdensome. Some vehicles will be adding streets to their route map literally ever day.

While I have been a professional privacy advocate, and I do care about just how the privacy of car users is protected, I am frankly not that concerned during the pilot project phase about how well this is done. I do want a good regime — and even the ability to do anonymous taxi — so it’s perhaps not too bad to think about these things now, but I suspect these regulations will be fairly meaningless unless written in consultation with independent privacy advocates. The hard reality is that during the test phase, even a privacy advocate has to admit that the cars will need to make very extensive recordings of everything they can, so that any problems encountered can be studied and fixed and placed into the test suite.

50 state laws

NHTSA’s plan has been partially endorsed by the self-driving coalition for safer streets (whose members include big players Ford, Google, Volvo, Uber and Lyft.) They like the fact that it has guidance for states on how to write their regulations, fearing that regulations may differ too much state to state. I have written that having 50 sets of rules may not be that bad an idea because jurisdictional competition can allow legal innovation and having software load new parameters as you drive over a border is not that hard.

In this document NHTSA asks the states to yield to the DOT on regulating robocar operation and performance. States should stick to registering cars, rules of the road, safety inspections and insurance. States will regulate human drivers as before, but the feds will regulate computer drivers.

States will still regulate testing, in theory, but the test cars must comply with the federal regulations.

New Authorities

A large part of the document just lists the legal justifications for NHTSA to regulate in this fashion and is primarily for policy wonks. Section 4, however, lists new authorities NHTSA is going to seek in order to do more regulation.

Some of the authorities they may see include:

  • Pre-market safety assurance: Defining testing tools and methods to be used before selling
  • Pre-market approval authority: Vendors would need approval from NHTSA before selling, rather than self-certifying compliance with the regulations
  • Hybrid approaches of pre-market approval and self-certification
  • Cease and desist authority: The ability to demand cars be taken off the road
  • Exemption authority: An ability to grant rue exemptions for testing
  • Post-sale authority to regulate software changes
  • Much more

Other quick notes:

  • NHTSA has abandoned their levels in favour of the SAE’s. The SAE’s were almost identical of course, with the addition of a “level 5” which is meaningless because it requires a vehicle that can drive literally everywhere, and there is not really a commercial reason to make a car at present that can do that.
  • NHTSA is now pushing the acronym “HAV” (highly automated vehicle) as yet another contender in the large sea of names people use for this technology. (Self-driving car, driverless car, autonomous vehicle, automated vehicle, robocar etc.)

This was my preliminary report. More analysis can be found under the NHTSA tag.

The incredible Cheapness of Being Parked

Some people have wondered about my forecast in the spreadsheet on Robotaxi economics about the very low parking costs I have predicted. I wrote about most of the reasons for this in my 2007 essay on Robocar Parking but let me expand and add some modern notes here.

The Glut of Parking

Today, researchers estimate there are between 3 and 8 parking spots for every car in the USA. The number 8 includes lots of barely used parking (all the shoulders of all the rural roads, for example) but the value of 3 is not unreasonable. Almost all working cars have a spot at their home base, and a spot at their common destination (the workplace.) There are then lots of other places (streets, retail lots, etc.) to find that 3rd spot. It’s probably an underestimate.

We can’t use all of these at once, but we’re going to get a great deal more efficient at it. Today, people must park within a short walk of their destination. Nobody wants to park a mile away. Parking lots, however, need to be sized for peak demand. Shopping malls are surrounded by parking that is only ever used during the Christmas shopping season. Robocars will “load balance” so that if one lot is full, a spot in an empty lot too far away is just fine.

Small size and Valet Density

When robocars need to park, they’ll do it like the best parking valets you’ve ever seen. They don’t even need to leave space for the valet to open the door to get out. (The best ones get close by getting out the window!) Because the cars can move in concert, a car at the back can get out almost as quickly as one at the front. No fancy communications network is needed; all you need is a simple rule that if you boxed somebody in, and they turn on their lights and move an inch towards you, you move an inch yourself (and so on with those who boxed you in) to clear a path. Already, you’ve got 1.5x to 2x the density of an ordinary lot.

I forecast that many robotaxis will be small, meant for 1-2 people. A car like that, 4’ by 12’ would occupy under 50 square feet of space. Today’s parking lots tend to allocate about 300 square feet per car. With these small cars you’re talking 4 to 6 times as many cars in the same space. You do need some spare space for moving around, but less than humans need.

When we’re talking about robotaxis, we’re talking about sharing. Much of the time robotaxis won’t park at all, they would be off to pick up their next passenger. A smaller fraction of them would be waiting/parked at any given time. My conservative prediction is that one robotaxi could replace 4 cars (some estimate up to 10 but they’re overdoing it.) So at a rough guess we replace 1,000 cars, 900 of which are parked, with 250 cars, only 150 of which are parked at slow times. (Almost none are parked during the busy times.)

Many more spaces available for use

Robocars don’t park, they “stand.” Which means we can let them wait all sorts of places we don’t let you park. In front of hydrants. In front of driveways. In driveways. A car in front of a hydrant should be gone at the first notification of a fire or sound of a siren. A car in front of your driveway should be gone the minute your garage opens or, if your phone signals your approach, before you get close to your house. Ideally, you won’t even know it was there. You can also explicitly rent out your driveway space for money if you wish it. (You could rent your garage too, but the rate might be so low you will prefer to use it to add a new room to your house unless you still own a car.)

In addition, at off-peak times (when less road capacity is needed) robocars can double park or triple park along the sides of roads. (Human cars would need to use only the curb spots, but the moment they put on their turn signal, a hole can clear through the robocars to let them out.)

So if we consider just these numbers — only 1/6 of the time spent parking and either 4 times the density in parking lots or 2-3 times the volume of non-lot parking (due to the 2 spots per car and loads of extra spots) we’re talking about a huge, massive, whopping glut of parking. Such a large glut that in time, a lot of this parking space very likely will be converted to other uses, slowly reducing the glut.

Ability to move in response to demand

To add to this glut, robocars can be the best parking customers you could ever imagine. If you own a parking lot, you might have sold the space at the back or top of your lot to the robocars — they will park in the unpopular more remote sections for a discount. The human driver customers will prefer those spots by the entrance. As your lot fills up, you can ask the robocars to leave, or pay more. If a high paying human driver appears at the entrance, you can tell the robocars you want their space, and off they can go to make room. Or they can look around on the market and discover they should just pay you more to keep the space. The lot owner is always making the most they can.

If robocars are electric, they should also be excellent visitors, making little noise and emitting no soot to dirty your walls. They will leave a tiny amount of rubber and that’s about it.

The “spot” market

All of this will be driven by what I give the ironic name of the “spot” market in parking. Such markets are already being built by start-ups for human drivers. In this market, space in lots would be offered and bid for like any other market. Durations will be negotiated, too. Cars could evaluate potential waiting places based on price and the time it will take to get there and park, as well as the time to get to their likely next pickup. A privately owned car might drive a few miles to a super cheap lot to wait 7 hours, but when it’s closer to quitting time, pay a premium (in competition with many others of course) to be close to their master.  read more »

Tesla Radar, MobilEye fight and the Comma One $1,000 add-on-box

Tesla’s spat with MobilEye reached a new pitch this week, and Tesla announced a new release of their autopilot and new plans. As reported here 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 and is central to the operation of the Telsa autopilot, they may have been surprised that MBLY stock took a big hit after that 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 intimating 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 are 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 of that and told Tesla that 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, and 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!

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 most 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 functionlity 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 (in particular the radar, since the Comma One has a 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 also that bolt-on products are also 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.

Syndicate content