Transferring Autonomous Driving Knowledge on Simulated and Real Intersections

We view intersection handling on autonomous vehicles as a reinforcement\nlearning problem, and study its behavior in a transfer learning setting. We\nshow that a network trained on one type of intersection generally is not able\nto generalize to other intersections. However, a network that is pre-trained on\none intersection and fine-tuned on another performs better on the new task\ncompared to training in isolation. This network also retains knowledge of the\nprior task, even though some forgetting occurs. Finally, we show that the\nbenefits of fine-tuning hold when transferring simulated intersection handling\nknowledge to a real autonomous vehicle.\n