China Autonomous – Dog and Pony Show.

Pony.ai & Baidu can do what Waymo apparently cannot.

• While Beijing is keeping the important autonomous driving data close to its chest, what is available suggests that Pony.ai is the company that is best placed globally to offer a solution in more than one country.
• Baidu is a close second.
• The Beijing Municipal Bureau of Economics and Information Technology has released some initial data on testing that was carried out on the streets of Beijing in 2018.
• Baidu has 45 vehicles and drove a total of 139,887km, Pony.ai had just 2 vehicles and drove 10,132km while NIO had one vehicle and drove 2,415km.
• Unfortunately, no disengagement data has been given and so it is impossible to determine who did well and who did not, but if I use the California DMV data (see here) as a guide, I can draw some very preliminary conclusions:
  • First, miles driven: is generally a good indicator of excellence in autonomous driving (see here) unless you are Apple.
  • Consequently, I would have expected Baidu to have reported the lowest number of disengagements per mile driven with Pony.ai coming in second.
  • The other data does not support this conclusion.
  • Second, California performance: indicates that while Baidu drove more miles in California during 2018, its vehicles disengaged 4.5x more frequently than those of Pony.ai.
  • Third, km per vehicle: Baidu had 45 vehicles which drove an average of 3,109km each while Pony.ai had just 2 vehicles and which drove an average of 5,066 each.
  • NIO also fares pretty well on this measure but there is no California data to compare NIO against its other Chinese rivals.
  • One explanation for the discrepancy could be that Baidu had more problems with its systems meaning that the vehicles spent more time being worked on rather than on the road driving.
• Putting this together, I would conclude that there is a pretty high probability that Pony.ai has outperformed Baidu in Beijing.
• There is a big caveat to this analysis which is that the deep learning systems that power these vehicles are very bad at switching geographies.
• Hence, I think that both Baidu and Pony.ai have trained algorithms from scratch in both geographies and that there is no ability to take the algorithm from one geography and use it in another.
• However, techniques for training, expertise at building models and other rules-based software would be transferable, and it may be these factors that are making the difference.
• Although Pony.ai would appear not to be the best at autonomous driving (ranking 4^th in RFM’s analysis), it appears to be the best non-North American offering and appears to be working in more than one geography.
• This is something Waymo needs to work on as taking its vehicles and running them in Arizona caused real problems implying that Waymo is fantastic as long as one stays in California.
• Consequently, although Waymo could theoretically drive from New York to San Francisco without a disengagement, I think it will be a very long time before this feat is attempted.
• To be able to offer a credible solution to a large vehicle maker one needs to be able to offer the ability to work in a range of geographies, and in that regard, the Chinese look to be ahead.
• Waymo, Cruise and Zoox remain the top three in autonomous driving, but there are signs that this will soon be challenged, especially in non-North American locations.