Autonomous Autos 2020 – The grind

Still a very long way to go.

• The lead changes once again as Baidu does not feature in this year’s data, but it is now very tight at the top and many players have seen little improvement in the last 12 months leaving us still very far from real commercial deployment.
• The latest disengagement report from the California Department of Motor Vehicles has been released and can be found here.
• The use of disengagements to measure the quality of an autonomous driving system is deeply flawed but in the absence of any other measure, it at least gives an idea of where the field lies.
• Furthermore, for the last 4 years, it has provided a reasonable approximation of reality backed up by RFM’s qualitative research in this area.
• Hence, I am comfortable with drawing some broad conclusions from the data.
  • First, tightening continues: There is now almost no difference between the top three players who are now Waymo, Cruise and AutoX.
  • All three of these companies are capable of driving between 20,000 and 30,000 miles without having a single disengagement.
  • This further underpins my long-term view that the market to supply self-driving software will be very competitive and that prices will fall very quickly once commerciality is reached.
  • However, this date is still pretty far away, and I see no reason to change the long-term forecast of 2028 for commercial autonomous driving.
  • Second, data reliability is falling fast: The total number of miles driven by many players is less than it was last year.
  • For example, Waymo drove 628,000 miles in California in 2020 compared to 1.5m in 2019.
  • This is because Waymo and many others are also testing in regions where there is no requirement to publish data.
  • This is particularly relevant in China where 5 of the top 12 companies testing in California in 2019 were Chinese.
  • Even with Baidu absent this year (see below), 5 of the top 10 in 2020 were Chinese.
  • This combined with testing going in many other states accounts for the falling total number of miles driven and the reliability of this data as a reflection of reality.
  • Furthermore, there is still no data from Yandex, Mobileye and now Baidu all of whom I think have strong offerings.
  • Hence, I am forced to use this data with increasing caution.
  • Third, Baidu is missing: Baidu chose not to test in California during 2020 despite being the leader in 2019.
  • I suspect that this is a reflection of the increasing tension between the USA and China where Baidu is one of the leading technology companies.
  • RFM rates its AI as global No. 2.
  • I think it likely that Baidu has moved a lot of its testing back to China and is focusing more on the home market given the possibility that the geopolitical situation may make it more difficult for Baidu to deploy in the USA.
  • There is no shortage of Chinese companies still testing but these are largely unknown and small start-ups and hence may be capable of flying under the radar.
  • Fourth, improvements continue to slow: The fact that there is now very little to separate the leaders is an indication that the rate of improvement is slowing.
  • There are also fewer companies testing and producing data which is both a sign of consolidation and the fact that testing is becoming far more widespread geographically.
  • Zoox and Apple are two that have not made much headway in the last 12 months.
  • Fifth, Apple is experimenting, nothing more: Despite the new-found excitement surrounding Apple’s apparent desire to make an electric vehicle (see here), its progress in autonomy has been very slow in the last 12 months.
  • In 2020 Apple managed to drive on average 145 miles without encountering a disengagement.
  • This is better than many of its competitors but represents only a marginal improvement on the 118 miles in managed in 2019.
  • This leads me to continue to believe that Apple is not taking this very seriously and is instead experimenting in this sector which may end up being more about making its own map as opposed to producing a working solution for an autonomous vehicle.
• The net result from this data is that there is now very little to separate the leading solutions and that this data is increasingly only capturing a portion of the real picture.
• This points towards the ongoing and continuing slowdown in development as improvements become harder and harder to come by given the limitations of deep learning for this use case.
• Furthermore, with little to distinguish the top offerings, rapid commoditisation is increasingly likely.
• This is pretty in much in line with RFM’s view that the sector is extremely overvalued (see here) and its long-held position of 2028 being the date when autonomy can go properly commercial.
• I continue to see no reason at all to any put money into this sector as there is already a massive oversupply of both capital and market participants.
• Furthermore, the valuations demand so much that only heavy losses beckon, even for investing in the winners.
• Rational humans should still steer well clear of this segment.