Also an interesting perspective from Jim Keller, that autonomy is easier than you might believe:
I think it was in this link but not 100% sure... I watched a bunch of his interviews.
Generally speaking, I think GPT3 shows the power of raw computer in ML: https://www.gwern.net/newsletter/2020/05
Cars in comparison can start at any moment, from any location and want to go anywhere they like without knowing what other cars are doing, or -worse- the road conditions (potholes, ice etc).
Putting blame is also easier, because legally all pilots agree to be recorded while flying since they're doing business. It'll be much harder I guess to allow recording inside cars.
The closest thing to that are collision avoidance systems that instruct pilots to go up/down by talking to one another and agreeing on a plan for a particular situation.
There is always a pilot communicating with ATC, providing data, receiving and validating instructions and then entering them to the flight computer.
The entire current approach has no correctness guarantees, ML systems are horrible at dealing with tail-risk, there's a human communication element to driving and so on. It's a ridiculously hard problem, honestly AGI hard.
I much more suspect that before we get autonomous driving we'll get more 'rail like' street systems, geofencing etc.. that limit how drivers and cars can behave, that's kind of how the whole hyperloop thing works ironically enough. Obviously if we wanted that we could just build better metro systems to begin with which already drive themselves.