So the training car has a max speed of 10 meters per second or 22mph which the net matches. Future datasets will contain the road's speed limit as well as throttle / brake targets required to reach that speed limit.
How much real-time is needed to get 1k (or 1M) iterations, to improve it so that it moves as fast as the speed limit allows while obeying all the laws?
Convergence on target speeds and steering is actually very quick (a matter of hours on a modern GPU / AlexNet / 227x227 RGB) once you have the data. Generating training data with the correct speed limit and providing speed limit based rewards is the missing step. The training car already obeys traffic laws.