Autonomous driving – Common roadblock.

Pony’s pilot highlights everyone’s problems.

• The launch of Pony.ai’s 1^st live pilot looks great but digging into some of the details reveals that, just like everyone else, Pony.ai is plagued with the limitations of AI which are going to take a long time to solve.
• Pony.ai has announced PonyPilot which is a pilot “autonomous” ride-hailing service available in Guangzhou that is available to Pony.ai employees and selected affiliates.
• This was announced at the Shanghai Auto Show and from the public announcements and materials, the following can be observed:
  • First, Geofencing: The service will be limited to a 50km² area within Guangzhou which has presumably been mapped in extreme detail and which is being constantly monitored for changes.
  • The high definition map has become a critical piece of the autonomous driving solution as the deep learning algorithms are really struggling to deal with infinite and uncertain environments (see here) such as roads.
  • A highly detailed map within a limited specific area creates a finite environment and reduces uncertainty that can allow the algorithm to function to a reasonable level of quality.
  • As the area in which the service is made available increases, the probability of uncertainty rapidly increases to the point where the algorithm catastrophically breaks down.
  • This is why autonomous driving demos are strictly geofenced and why they often quickly run into trouble outside of these areas.
  • Tesla’s Autopilot is not geofenced (see here) but I consider this to be ADAS rather than autonomous driving.
  • Second, safety drivers: If the videos are correct, then the pilot service will be sporting not one, but two safety drivers in each vehicle.
  • In my opinion, a service that uses a safety driver is not an autonomous offering as the whole point of having a vehicle that pilots itself has been obviated.
  • When it comes to safety drivers, the best demo I have seen is Yandex which had one safety driver in the passenger seat and no one in the driving seat (see here).
  • However, Yandex’s demo was carried out under strictly controlled and geofenced conditions with a set route although it was on real roads with real traffic.
  • Two safety drivers implies that the system is expected to disengage pretty often meaning that a lot of work is still needed to get it to commercial reality.
  • Third, fixed starts and stops: Although the area covered is only 50km², the pick-up and drop-points have been predetermined (see here, paragraph 5).
  • This means that all of the available routes can be worked out ahead of time and programmed into the system.
  • This is yet another method of dealing with the uncertainty that an autonomous system is being asked to deal with so that it can function to an acceptable level of performance.
• Based on the California (see here) and the Beijing data (see here), Pony.ai fares pretty well against Western competitors and even against Baidu, but everyone is struggling with these issues.
• The net result is that a different approach to autonomous driving is likely to be needed and rethinking the system again from scratch is going to take time.
• I am sticking with my 2028 commercial availability for autonomous vehicles for the moment, but it is starting to look a little precarious.