Re: Missing the obvious
Assuming a test environment I have seen is indicative of the general case, the AI is training on conditions recorded from the road. This is video and depth information gained from sensors similar to those on a autonomous vehicle from an urban environment. It has poor road markings, shadowed areas from trees, etc.
There are weaknesses, for example all footage is in good weather, sun above the yardarm, driving on the right, traffic patterns are the same at each location; assuming the machine learning is as adaptive as AlphaZero, then different AIs will familiarize with local conditions so training for different legislative environments will be necessary.
I expect collecting large data sets from roads will be an early milestone for any serious automotive vehicle developer (or a supplier company will be created with a super-Streetview dataset). It would be interesting to see this data modeled and then a large number of AIs being run in it simultaneously in a MMORPG mode so they have to navigate around one another.