Curious; If you train a neural net to drive a car with X number of samples, and the car does something erratic, how do you determine which data item introduced the erratic behavior and how do you go about untraining that behavior?
Whatever the sensing technology, everyone uses it to build a "state of the world" model which is then fed into algorithms for planning movement. Nobody (to my knowledge) just wires up input pixel neurons to output pedal neurons.