Arguably it's not a pure ML problem where you can just learn what to do from human drivers. You need to know what the alternatives are, which actions are potent

Arguably it's not a pure ML problem where you can just learn what to do from human drivers. You need to know what the alternatives are, which actions are potentially dangerous etc. So you run your "beta" system and record when the safety driver takes over. Then you improve the system from that and drive some more. These iterations will take a lot of time, and you can't just scale it with more engineers (see The Mythical Man Month etc).