Check out our new work on learning to drive in new cities without collecting any new demonstrations in that city. All we need is a simple reward function and readily available map and meta-information about the city. Self-play reinforcement learning does the rest!
