>but have the computers learning through that constant, immediate feedback.

I don't know what is stopping them from simulating everything inside a computer.

Record the input from all the sensors when a car equipped sensors is driven through real roads by a human driver. Replay the sensor input, with enough random variations and let the algorithms train on it.

Continue adding to the library of sensor data by making the sensor car by driving it through more and more real life roads and in real life situations. Keep feeding the ever increasing library of sensor data to the algorithm, safely inside a computer.