I don't believe that you can use traditional algorithms to teach the car street driving, because there are to many different variations... Even if your autopilot is 99% correct and you drive 20000km a year, you still drive wrong 200km of it.

Exactly!

And this is why, if the problem is solveable, it must be solved by learning models shepherded by expert engineers. The LLMs can take care of the long boring stretches, freeing skilled engineer time to fine-tune an LLM algorithm hybrid for the tricky bits.

I'm inclined to believe the problem is solveable, but since I'm not selling anything, I'm allowed to say "if". Heh.