Why would this be different with a net that only handles the input? It would be the same. All v12 does is learn control outputs directly from camera data instea

Why would this be different with a net that only handles the input? It would be the same. All v12 does is learn control outputs directly from camera data instead of, for example, learning to detect a traffic light and then have some hand rolled logic. If the input is unpredictably new it would have failed in either case.