Autonomous Driving – Back in business?

Cruise makes a cautious return. 

• It looks like Cruise is about to resume testing its robotaxis service which makes sense because the penalties that it incurred were more a result of bad behaviour by Cruise executives as opposed to the technology being more faulty than anyone else’s.
• Cruise has been off the streets since November 2023 as a result of an incident where a human-driven vehicle propelled a pedestrian underneath a Cruise vehicle which then dragged the pedestrian 20 feet as it moved to the side of the road causing significant injury.
• As usual in these situations, this was a situation that the AI had not been trained for meaning that the vehicle’s actions caused the injury of the pedestrian.
• This is not a problem that is specific to Cruise but an issue for all of the autonomous driving industry that is using deep learning systems to interpret the world perceived by cameras and lidars into a representation of reality.
• Systems based on deep learning cannot deal with a situation in that they have not been specifically taught meaning that when the task at hand is almost infinite and highly variable (like the road), they really struggle to perform to an acceptable level of reliability in real-world conditions.
• In the USA, human-piloted vehicles kill 1 person for every 97m miles driven which is the standard that machines need to significantly beat if they are ever to be properly commercialised.
• No one, not even leaders like Waymo, Baidu, Cruise or Mobileye, is close to this standard which tells me that there is still a very long way to go.
• This combined with a starvation of capital from the sector now that autonomous driving is deeply unpopular, means that it is likely to be many years before this is a reality.
• However, this was not Cruise’s problem and in my initial assessment of this incident, it is clear that I was not cynical enough (see here).
• The California DMV that overseas autonomous testing in California accused Cruise of deliberately omitting video evidence from its incident report which I put down to a misunderstanding.
• Turns out that I was wrong and in fact, Cruise had wilfully misled the regulator which resulted in its licence being suspended as well as the firing of the executives involved.
• However, it looks like Arizona is willing to consider letting Cruise back on its roads which makes sense as Cruise’s problems are more about its behaviour than its technology which is an easier fix.
• I suspect that if the incident had involved Waymo, Mobileye or Baidu, the outcome of the incident would have been similar although I doubt that they would have been foolhardy enough to mislead the regulator.
• Cruise and I suspect everyone else will now be more cautious than ever which I think is likely to slow development further.
• One idea to solve this problem is to train the algorithms in simulations because then situations that cannot be recreated in real life can addressed with no ill effects.,
• However, while simulation has its uses, I do not think it is any good for training algorithms to deal with real edge cases because only reality can create them.
• Using generative AI is also unlikely to be of much use because I think that generative AI can only recreate and remix what it has already seen and so can never really be creative meaning no genuine edge cases are created.
• Instead, I have long believed that the solution lies in addressing the real problem (lack of causal understanding) rather than trying to mitigate its side effects.
• Progress on this front is slow and I am in real danger of being forced to revise my target for autonomous driving which has for years been 2028.
• Cruise now has a shot to get back into the race, but I am not 100% convinced that the race is worth running.
• The delays have meant that all of the stragglers who had enough money have had time to catch up and now there is not that much to differentiate between the many leading offerings.
• This means that when this does come to market, there will be many offerings and that robotaxis will not deliver the fat 66% gross margins that Mr Musk is predicting but will be a bloodbath of cutthroat competition.
• I suspect that OEMs will have several equally good solutions to choose from and so I am dubious as to whether it makes financial sense to develop in-house as it is unlikely to be a big differentiator.
• This is why I remain cautious about the financial outlook of autonomous driving and robotaxis in general and would instead look to invest in companies that benefit from growth in greater autonomy.
• This is why I own Ouster (digital lidar) which has been on a rough road but the company has prevailed and is in a good position to grow its way to profitability without having to sell a single lidar to a car company or raise any more money.