Anthony Levandowski, Jiajun Zhu and Dave Furguson, all formerly of Google Car, make big announcements


By coincidence we see two significant announcements today from people who were former leaders on the Google car project, now in their own companies.

A Nuro drives on a public street

The more significant is Nuro's automated delivery service which uses their unmanned road robots. The post claims it is the first unmanned delivery service, which is quite false, since robots from Starship (in which I am involved) as well as competitors like Kiwi, Robby, Marble and others have been running delivery service on sidewalks for some time now.

However, it is the first unmanned road based service of any kind, pending some details, which is much more significant. However, the vehicles have both remote operation and, for now, they are followed by chase cars which have some ability (not yet disclosed) to take over the robot or cause it to shut down. Depending on the quality of the radio link to the robot, that's like having a safety driver -- you could even give the passenger in the chase car a VR headset so it would seem like they were sitting in the robot, if it would not make them sick.

Deliverbots present an easier problem than passenger robocars. They can limit themselves to easier and slower roads. Their cargo is no anxious or in the type of hurry humans get into. You can't kill your cargo, though you must still never hit other people. You can limit your travels to where radio signals are good so that remote operators can always solve problems. No signal down that street? Just don't go down that street, at least for now.

That's why we can see this first.

Anthony Levandowski returns

I know Anthony well, so I don't comment about him a great deal. I am both surprised an d unsurprised to see him announce He's back with a truck co-pilot with more details in a Guardian article about a cross-country trip with no disengagements using no LIDAR.

Anthony devoted 6 years of his efforts to LIDAR, and his efforts to create a LIDAR at Uber via Otto were the subject of a famous lawsuit that brought down him, Otto and parts of Uber. Today he reverses himself and declares, like Elon Musk, that LIDAR is indeed a crutch. He's building a $5,000 "co-pilot" system to sell to truckers to make their job easier, and doing almost all of it with neural networks.

Anthony faces an unusual problem. In a lot of the industry, his name is poison because of the past events, whether allegations are true or not. This makes it unlikely a company he leads could sell to major OEMs or get investment from major VCs. However, he can sell to individuals, such as truck drivers.

There are many small startups trying to make the AI-not-LIDAR approach work. Whether it is good to make an autopilot that way depends on your view of autopilots. From a technological standpoint, they can do it. The harder questions revolve around what the human does in concert with the tool -- do they become complacent or stay safe? Do their driving patterns change in a way that is good or bad? Tesla says they get better, but the jury is still out. My anecdotal experience is that after driving with autopilot (or even adaptive cruise control) for a while, when I switch off, I need to consciously remember that I am now in charge of the brakes. The wheel is not a problem. But we've had ACC for a long time without much trouble.


The 25 mph max speed is my biggest worry. No doubt it is temporary, but in the meantime they have vehicles which will no doubt cause an inconvenience to other road users. In order to get to the next level they probably need lots of road miles. The more they need, the more annoying they will be. I understand they will drive in convoy with a support vehicle with remote control capability. Perhaps given the low mass of the vehicle they would be better to go at the speed limit (35 mph?).
BTW if you have any updates on the starship buggy you are at liberty to share, that would get my vote.

I am afraid the law may or may not restrict them to 25mph. This is an area of flux. In theory any "vehicle" has to follow the FMVSS, and that's super expensive, so there is the NEV exception which makes it much more possible to build a vehicle, but it has to be limited to 25mph. However, there are so many tweaks to this:

  • The FMVSS is mostly about protecting passengers. The Nuro has none. Perhaps it doesn't even apply
  • While it may die this congress, laws are coming which allow large numbers of exceptions to the FMVSS per company
  • There are arguments that the FMVSS applies only to vehicles you sell not vehicles you own yourself and operate. However, I don't know how long that loophole will last.

Starship is avoiding all that by not being a road vehicle. Of course that makes it slow. Starship just started a program in a few places offering last mile service so people can be sure their packages are delivered only when they are home.

Neither of these stories impress me. Co-pilot has the problem of all these advanced driver assist technologies. It is simply unrealistic to pretend that a human will stay alert over stultifying hours of nothing to do, ready to instant spring into action when some unspecified circumstance requiring this should arise. It is a great way to blame the user when disaster strikes, but it is a terrible way to design a safe system.

As for the delivery system? As currently implemented, this driverless system requires two vehicles, and apparently two humans. Presumably this is intended as an intermediate stage, but it nonetheless is something much less impressive than the hype, including that it can only sort-of barely handle the mean streets of Scottsdale--at least the easier ones.

Yes, if you were looking for final products, everybody "disappointed" this year. If you are looking for milestones on the way, then Nuro's project is one of them. Not long ago a truly unmanned (in fact incapable of carrying people) vehicle on public streets would have been viewed as science fiction. Chase car or not.

Anthony's product is indeed less of a technological milestone. A full cross country with no interventions also would have been science fiction not too long ago. Doing it with a small team is now on the "not world shaking" level but still a first.

I am reacting based on the prior claims by the tech companies. If the defense is that these were an obvious pack of lies from the start, I agree. That's what I have been saying all along. If this is followed by "but now they are telling the truth!" then I will roll my eyes. My reaction when Google cars roaming the streets of Mountain View were first being hyped was that the discussions had the air of "We have solved 95% of the problem. How hard can that last 5% be?" heh. Now we have progressed to "We have solved 96% of the problem..."

Waymo has always been quite understated in their claims. They are not changing their story, other than the story of when they can go live.

How impressive the delivery system is depends on how often humans need to take over. If the humans are just there because it's unethical to let the machines loose until they have more driving time, it's pretty impressive (though maybe it wouldn't be if I knew more about the state of the art). If, like the other announcement, the system is known to be unable to handle certain situations that might come up (and the standby driver is not easily familiar with how to recognize those situations), my feelings are that it's unethical to deploy it on public roads in the first place (like the other announcement, like Uber, and probably like Tesla though I'm not completely familiar with what situations Tesla can't handle).

There are two kinds of "situations." One are low speed situations, with strategic help needed. The chase car will probably not be involved in that. The car will pull over or stop quickly in such situations and the remote control center -- which never goes away even in full deployment -- will instruct the robot on what to do.

The other kind of situation is tactical, emergency decisions to stop or swerve. That's what the chase car safety driver would do, and that should be very rare.

Add new comment