Robocar 2023 In Review: The Fall Of Cruise

Looking at 2023 in review for self-driving cars, one story stands out as the clear top story of the year, namely the fall of General Motors’ “Cruise” robotaxi division, which led to a pause of operations for Cruise over the whole USA, a ban on operations in California, the resignation of the founder/CEO and much more. That was actually the most prominent component of the other big story of the year, namely the battle between San Francisco and both Cruise and Waymo.

Cruise and Waymo are, by a large margin, the leading robocar companies in the west, and arguably the world, though the Chinese might argue with that. Both are based in the San Francisco Bay Area and worked to do early pilot deployments in the city, both because it’s home, but also because it’s both complex enough to offer most of the challenges you need to overcome, but leaves out a few, such as snow, that can be left for later. But though San Francisco may be seen as the world capital of technology, it’s also a luddite capital, having been one of the few places to ban the Segway and delivery robots before they even got started.

San Francisco had little say in the deployment of the robotaxis. Driving on the streets is regulated by the state of California, not individual cities, and for good reason. The Department of Motor Vehicles licenses cars and handles rules of the road, and the Public Utilities Commission regulates taxi services. San Francisco didn’t like that at all, and officials there felt the companies didn’t give them proper attention.

I go into the details of the flight that took place between various city officials and the Robotaxi companies elsewhere. It was a strange fight, and a harbinger for the introduction of the vehicles in some areas, and general public reaction to AI which has big real world consequences, both good and bad. While the city does not have regulatory authority over the companies, it used the powers it had to create issues and to lobby the state agencies which did have power, and that probably played a role in Cruise’s downfall.

The Pedestrian Impact

The climax of the fall of Cruise was a serious injury crash started by an unknown driver, who hit a pedestrian crossing against the light. This driver fled the scene. The pedestrian went up on the hood of the Nissan, then slid off to the right, in front of the Cruise Bolt which was driving in the right hand lane. The Cruise detected her and hit the brakes, but still ran over her. Up to this point, it would be hard to fault the Cruise for any of this, which was the fault of the Nissan driver, and to some extent the pedestrian, in a crosswalk but crossing against a Red/Don’t Walk sign.

What happened next changed everything. After the Cruise had come to a stop, it quickly followed programming to try to pull off the road. It couldn’t really, as the bike lane to the right had bollards isolating it, but it pulled forward about 20 feet and slightly to the right. The pedestrian, however, was under the car, and was dragged this 20 feet at up to 7 mph. When the vehicle finally stopped, its rear wheel had rolled up on her legs. She was injured and in pain, but procedures here–reaffirmed by emergency crews when they arrived, are not to try to drive off of somebody. Instead, they waited until they could bring in the “Jaws of life” to jack up the vehicle and extract her from under it. She went to hospital where she is slowly recovering, months later.

The Coverup

Even this serious mistake should not have and perhaps would not have driven Cruise to its fate. The bigger mistake was they did not want to talk about it. They were eager to show the car’s recorded video of events to the press, and to officials, but only over Zoom. They did not mention the dragging, even when asked about it. I did not ask to see the part of the video after the crash, but others who did see it say it ended before the dragging. The DMV states they were not shown the dragging, and Cruise did not tell them about it, including in a letter they sent shortly after the events. Cruise insists they did show the full video to the DMV, and the DMV insists otherwise, but there is little doubt that they were not open about what was obviously the most important part of the chain of events when it came to understanding the role of the robotaxi in the calamity. Cruise was very eager to show that the initial crash was not their fault, but didn’t want to talk at all about their own serious mistake.

As is often the case, the coverup is worse than the actual event. A few days after this event, the DMV, which had been investigating other Cruise events–in part due to Cruise’s troubles with San Francisco–sent an order to Cruise to cease operations in California. The order cited two core reasons: That the vehicles were not safe, and that they had deceived the DMV about it. While they have made no statement about which of these two was the most important trigger, there is some suspicion that it was the cover-up that led to such a stern order. After this order, the PUC also pulled Cruise’s permits, and also cited the cover-up, even citing the later claim that they had really shown the DMV the full video. The regulators were very upset with Cruise’s lack of candor.

It’s definitely bad to drag a pedestrian, and I’ll go into some theories as to why that happened. At the same time, anybody looking at robotaxis, including regulators, has to know that they are not and will never be perfect. There will be serious incidents if they drive for long enough. For now, everybody is asking if these serious incidents are happening at an unacceptable level, which is to say, not sufficiently better than happens with human drivers. While we might argue about it, we do accept the risk rates of human drivers, even though they are high. Robotaxis are promising to do better than that, and lower risk on the roads. If they give evidence that they are doing that, one will have to tolerate some incidents in order to reach the future goal of a large deployment that reduces risk a lot.

As such, the dragging on its own should not have led the DMV to pull Cruise’s permits, at least not immediately. If they found, through their study of all Cruise events, that their record was unacceptably poor, it might be the straw that breaks the camel’s back, but Cruise had been publishing numbers saying their record was good and within suitable levels. When we add the cover-up, we might decide not to trust Cruise’s reports, which complicates this question.

Dragging a pedestrian is a very special class of robotaxi accident, because it’s horrifying. All pedestrian injuries are horrifying, but in particular the car’s actions seem especially “inhuman.” We have a deep fear of the inhuman and uncaring robot harming a human, oblivious to what it’s doing. That’s likely the case here–the Cruise car almost certainly was not aware the pedestrian was underneath. While an injury is an injury, public trust, fear and perception matter. As such, I have recommended since the start of these projects that they make special efforts to avoid this sort of incident. Indeed, by strange coincident in only discussion just a few days before this incident, I warned that dragging a pedestrian would be a nightmare situation, and teams must work to avoid it. Cruise didn’t do enough, and while I had advised them in the past, it was not extensive enough to get into this level of recommendations. Had they done better there, the rest of the story might never have come to pass.

Cruise may also have not sufficiently studied situations like a pedestrian being thrown in their path by another car. All teams study pedestrian impacts, both in simulation and by using test dummies as their test tracks. This particular situation is rare enough it may not have been tested much, if at all. As noted, anything in a situation like that would not generally be their fault. While robotaxis want to avoid all bad outcomes, they will naturally focus on ones for which they would be blamed, as the law strongly sets incentives that way.

There is a highly ironic possibility. Cruise had been taking a lot of criticism for blocking streets when the cars got confused by the situation. In particular, in cases where their remote operations team–who very frequently are resolving problems for cars–can’t confidently do their work, they will even dispatch a rescue driver to rush over to the car and drive it away. The problem with that is that while it is conservative and safe, it blocks traffic for 5 to 15 minutes and nobody is happy. In theory, somebody should be happy because we want these pilot project vehicles to be conservative. Better to block traffic than create safety risk. It’s a temporary state. But those blocked–particularly the fire department, do not see it that way.

It’s possible that Cruise took actions to have their cars clear scenes and pull over more quickly. In this case, it seems the car went to pull over almost immediately after stopping, where the better course would be to bring in a team of many remote operators to review all the video and decide action. Those remote operators would have noticed the pedestrian went under the car and didn’t re-emerge and immediately ordered it not move. The car should have noticed this, which is its error, but the humans would do better. We don’t know this is the case, but it’s been revealed that remote operator evaluation is very frequent for Cruise, so it seems almost impossible to imagine that an impact with a pedestrian would not immediately call for extensive remote human analysis–yet it didn’t, and the car pulled over on its own, dragging the victim. As such, I suspect this was either a result of pressure to clear the road quickly, or a deeper flaw in how they modeled the situation.

When it comes to the cover-up, this was sadly not too surprising consider Cruise’s past patterns of communication with the press. I worked extensively with Cruise’s press office over the years and regularly criticized them for keeping things too close to the vest. This criticism was not simply because I wanted to know more, I advised them many, many times that they were acting against their own interests in doing so. The junior PR staff did not have authority to change policy. In almost all cases, they would issue a prepared statement and it was pulling teeth to get them to elaborate. This pattern of never volunteering information may have led them to keep mum about having dragged the pedestrian. It’s less likely there was a deliberate decision to cover that up, as anybody thinking about it for long would realize it would eventually come out. The victim was still alive and there were other witnesses, and detailed 3-D logs were stored inside the car which regulators and courts could access. And indeed, as cover-ups go, it was a very short one, lasting just a few days.

Waymo is not perfect either, and they often keep to the official statement. But they have been much more forthcoming in publishing safety results, though what they publish always makes them look good, which naturally raises questions. Waymo published details on every real and simulated “contact” event in their operations near Phoenix, and I told Cruise they need to at least do that, but they would not. Cruise would use phrases like “We have never had a life-threatening injury” which strongly says to me that they have had a near life-threatening injury, but they would not talk about it. In all cases, riders in cars involved in crashes are not available for interview, and do not post to social media, which strongly suggests they received settlement payments which include a non-disclosure clause. We see tweets and reports regularly about minor incidents with the cars from bystanders and occupants.

It is not hard to understand why the companies want to keep it close to the vest. This is a challenging project, and as noted, the public is not good at understanding that there will be mistakes during the development and testing. Nobody wants to test what triggers fear and distrust in the ordinary public who don’t study the complex issues of this. But even though the public may have trouble handling the truth, hiding the truth won’t work in the long run. Instead, one must find a way to help the public understand the truth and why the end result will be safer roads for all, or to help regulators make rational decisions if that’s not the end result.

The Cruise event may change some attitudes on that. I believe teams want, for every incident, to help people understand why it happened, but more importantly why it won’t happen again. The great benefit of robots is if they make mistakes, you can reprogram them to not repeat those mistakes. Humans aren’t like that. If one human makes an unsafe turn, even if they get a ticket or die, it won’t stop others from doing the same. Not so with robots. We put student drivers out on the road to make mistakes and learn, but only that one driver learns, not a whole fleet.

In future, robotaxi companies may need to follow a plan with much greater transparency than they might enjoy. This will disclose a few proprietary secrets, and expose some embarrassing ones, but there may be no other choice. Once you are deployed on city streets, your mistakes will not stay secret for long. It is vital that regulators and the public look at data, not incidents and anecdotes. The real story must be the story, not the story on social media.

In addition, the usual method of regulating in the USA is self-certification. That seems crazy to some, but the reality is that in a field of rapidly changing technology, where new techniques are invented regularly, it’s almost impossible for governments or 3rd parties to test and certify the freshly minted approaches. 3rd parties have to stick to conventional wisdom and they probably won’t understand the innovation and will almost surely impede it. But for self-certification to work, the companies have to tell the truth. Their first motivation for that are the harsh penalties they will get for not doing so. Cruise hid the truth and suffered a near-mortal blow, but for unknown reasons the fear of that didn’t stop them from hiding the truth.

Did Cruise push too hard, to the point of being unsafe?

In one sense, there is a clear answer to whether Cruise pushed too hard, in that the DMV felt so and pulled their permits. As yet, we don’t know if their incidents were enough to get that result, or if the cover-up was the key factor. Whatever the answer, there was definitely an impression that Cruise was pushing hard at growth, even harder than Waymo though there is general agreement that Waymo’s technology is superior. (That’s not too bold a claim considering that Waymo has been at it for several years longer, and was drawn from the unmatched technical teams at Google X, and they met all milestones sooner and had fewer significant incidents.) Waymo limited itself to just Phoenix and San Francisco, while Cruise was deployed in a few more cities and had announced it would deploy in 10 cities very soon.

Further evidence comes from leaks which suggest that Kyle Vogt, CEO and founder of Cruise, was pushing hard for that expansion and was receiving pressure to deliver results for GM. The leaks suggested Vogt had loosened certain safety criteria to allow the expansion to move more quickly.

There is also something to be learned from the pattern of incidents Cruise had. While we didn’t get full details on many of them, several of them suggested “rough edges” which should not be present in a live deployment. In one incident, a Cruise vehicle hit the back of a bus (nobody hurt except the bumper of the Cruise car) because of an obscure bug in their perception system. They fixed that bug quickly, but the deeper issue was their car didn’t have backup systems which would prevent it from running into a fat target like a bus no matter what perception heuristics were saying. It should be almost impossible for the car to hit a bus, except deliberately, no matter what goes wrong with the high-level perception, because much simpler systems can detect that such a crash is going to happen with 100% reliability. A bus is, after all, “the size of a bus” and you can’t miss it even with the simplest of sensors.

A few other incidents such as a left turn from a center lane suggested something more seriously awry. Ditto the car being unable to see caution tape and downed power lines–these are common things which should have been well tested in simulator. So is a fire truck moving into the opposite lanes to get around traffic stopped at a red light. A good self-driving team will have made simulation tests for all the things that anybody has heard of or encountered or thought of. It won’t be everything, but if it’s on your own list of “I’ve seen that” then it’s pretty common and should be in the test suite. Cruise gave off strong signals that their test suite was not as complete as it should be, or was not being applied very well. We can’t be sure of this, as we need to know what actually caused each incident to judge this, but the pattern is indicative.

Competition

If Cruise rushed more than they should, it was probably in fear of Waymo. Waymo, as discussed, has a commanding lead. They started year ahead of all other significant teams. They recruited most of the best people–indeed Waymo’s past team members (which include myself) have founded or run companies like Aurora, Argo, Kodiak, Nuro, Otto and Woven Planet, and been key members of most of the rest. Waymo has the resources of Google, easily one of the top tech companies and brands in the world, and owner of the Android platform. The Google family has immense compute resources and for a long time was the undisputed leader in AI technology, though that lead has been eroded. Waymo built its own LIDAR and radar but did not need to make cars or other hardware.

How do you compete with that? There were several angles Cruise may have looked at:

1. As a unit of GM, for all the downsides of being part of an old-world OEM, when it came time to make vehicles, particularly a custom vehicle, GM’s resources and skills are formidable. As such, Cruise put a lot of effort into the custom Origin vehicle. (Zoox also feels a custom vehicle will be its edge in the space.)
2. Cruise decided to develop and deploy its pilot in San Francisco, a much more challenging service area than Chandler, Arizona/Phoenix where Waymo did its pilot. Cruise believed that by starting with the hard problem it would be able to serve more complex cities faster, and cities like SF are real, demonstrably profitable markets, while Phoenix is less so. (Sidecar, Uber and Lyft and many others were born in SF.) To reduce that complexity, Cruise did its first pilot only at night.
3. Cruise pushed fast to expand, opening up in Chandler, Austin, Dallas and Houston, with plans to open in a total of 14 cities fairly soon, including in the UAE and Japan. Cruise did this in the belief that there is a literal first-mover advantage; if a company can establish in a city first, it can own the city, as Uber seemed to do over Lyft in so many cities.
4. Cruise even began advertising, sponsoring the San Francisco Giants’ uniform, even though demand was larger than the supply in the pilot services.
5. Cruise pushed hard to put more cars on the road and do more miles than Waymo was doing in San Francisco. That got them more of San Francisco’s ire, even if they had managed to match Waymo’s safety record.

Unfortunately, it’s not clear any of these was the right strategy. There is certainly debate about the value of custom vehicles. Waymo is putting in a semi-custom vehicle from Zeekr, but it’s not nearly as distinct as the Origin or the Zoox. The question of how strong the first-mover advantage is remains an open one. There are many historical examples of it working, and many of it not working. However, nobody thinks it’s a bad thing to be first as long as you’re good, so they do it. Apple (which is also competing in this space) is pretty famous for rarely being first, but coming later to dominate.

It’s not clear anybody else is nipping at the heels of Cruise. Motional, Zoox and Aurora are reputed to have good technology, but Aurora has done almost no car testing, and the others aren’t anywhere near where Cruise and Waymo are. Tesla has many fans but is not really even in the race, their technology being very far behind, though their assets as the world’s most innovative automaker with a large fleet of advanced cars are considerable. There are several serious Chinese players, but the current geopolitical situation makes it very unlikely they can compete in the USA or most of the west. The USA will not likely tolerate a large fleet of robotaxis controlled from China on US streets. Ford’s Argo was shut down and all the other auto OEMs have scaled back their projects or possibly hope to work with MobilEye. MobilEye has many assets but is also not nearly as far along in robocar deployment as Waymo and Cruise.

The Future

To understand the fall of Cruise, we ask the questions, “Why the secrecy?” and “Why the rush?” Some have speculated the rush may have been due to the pressures of a legacy publicly traded automaker like GM. Cruise’s previous CEO, Dan Ammann, was fired by GM CEO Mary Barra allegedly because he wanted Cruise to work only on its own robotaxi projects and to go public, while GM wanted more of what Cruise did to percolate up into GM. Cruise published very ambitious billion dollar revenue forecasts and was spending money at a rate that few parents would tolerate without a lot of confidence. There was probably pressure to provide proof that this revenue could actually be delivered. Alphabet has been, so far, a more patient parent, though Waymo has not been immune from layoffs. Both have been better parents than Ford.

It should be noted that if Cruise’s claimed safety record, superior to that of ridehail drivers, was true–we unfortunately must be skeptical of all their reports–then they weren’t moving too fast. Indeed, Waymo may be moving too slow. At least if the standard is “who is going to do the most to improve road safety?” because scaling first means reducing risk more and sooner. The main thing we can be more sure of is that Cruise tried to hard for a company that also lacked transparency, to the degree that it would eventually hide important things from regulators, so that the combination of crashes and secrecy would make the regulator conclude they were pushing too hard.

It is harder to find a good solution to the secrecy. Cruise’s example shows that companies with too much secrecy will be punished extremely hard if they slip up, so Cruise acted against its own interests. That it was doing so was clear to me, yet not clear to those on the inside.

Some would feel this means regulators must get inside the companies and dictate some of their procedures and safety methodologies. I believe this would be a major mistake. Their job is to innovate, and while they should learn from conventional wisdom and established best practices, they must not be bound to follow them. Instead, we need to work harder to make sure their interests are aligned with doing a good job and telling the truth about it, and if those drift, to fix it. The lesson of Cruise should actually do a lot to strengthen that message.

Cruise has fallen, but it’s not dead. It will cut 20% of staff, mostly those who managed operations which are now halted. GM’s tolerance for Cruise’s mistakes will be thin. Executives have been purged, and all will feel once burned.

Cruise may not return any time soon to San Francisco. Earlier, in spite of the troubles, this was unthinkable–they had put so much of their plan to beat Waymo on the fast they did SF first. Now the changes are so radical that this seems an easier decision. While they must satisfy the DMV some day, they may take their time, and go elsewhere. They are still planning projects overseas, though it is difficult to do your pilot across an ocean. SF has become so inhospitable that even Waymo may scale back there to focus on Los Angeles, Phoenix and Austin before seeking another city as complex as SF, even though SF is their hometown.

I hope the new Cruise will be more transparent, and not because they’ve been punished for failing in that area, but because they come to actually see the virtue of it. They have lost trust, and only a public which feels it understands the truth about their vehicles will give back that trust. Indeed, they have probably hurt the whole industry. I doubt that anybody else will start in San Francisco, even Zoox and Nuro which are based in the Bay Area.

Cruise’s new, slower plan will need the endorsement of GM management, so that if there was pressure to go faster, it is somewhat eased. It will never go entirely away, however, not when the budget is that high.

Cruise also should take this time to improve the technology. To avoid dragging anybody, they might put some ultrasonic sensors under their car, but they can probably also do that with better perception and object permanence software that understands when somebody may have gone under where they can’t be seen. I suspect they need to do a lot more in simulation. Both Cruise and Waymo have to work to avoid stalling in emergency scenes and avoiding the encounters before they happen, and they need to seriously reduce the need for rescue drivers which leaves vehicles stranded for many minutes. This is just the start of their todo list.

It is important they come back. While the smaller companies will eventually start real pilots, it is important to have two large competitors. As much as I like Waymo they should not be in this game alone. Yes, there is going to be a land rush as companies try to be the first mover in important cities, though in time I don’t see a repeat of the Uber/Lyft battle in the robotaxi world. I think if a new robotaxi comes to town with clear advantages they can become a contender or even take over. Lyft and Uber are close to identical, and so the more famous name with more cars is likely to win. Zoox, Waymo, Cruise and Motional won’t look the same.

It is, unfortunately possible that a firm like GM might get the same cold feet about robotaxi that shut down projects at Ford and many other OEMs. They may decide to put Cruise’s technology into a consumer car, one that can drive freeways and arterials for owners, but not offer a taxi service. There will be demand for such a product as a luxury feature, and that’s something GM understands well. The sensor suite is expensive but a cheaper one can do this job and will get cheaper still at GM volumes. The main challenge is that while freeways are much less complex than cities, their speeds mean that crashes can easily be very severe. That’s one reason Cruise and Waymo still don’t serve freeways even though they are crucial to serve large cities like Phoenix and LA. Luxury features are nice, but they don’t change the future of transportation. On the other hand it’s not clear if companies like GM want to see the future of transportation change. They’ve been doing fine for a century making cars and may want to not see that turned upside down. But if they don’t, somebody else will, and leave them behind.