When it comes to "closed track" and other "simulation" style things, you need to be very careful when ML is involved.

I don't know if Uber/Waymo/Others have "solved" this problem, but ML systems have a tendency to "learn" the wrong thing. You could run millions of simulations (both virtual or on a closed course) with "test obstacles", but the second the car encounters one a child with red balloons in the early morning on a foggy day it might not know what to do with it because that exact scenario wasn't in any simulations.