Learning to Drive from Observations while Staying Safe

The simulation of real-world traffic is a challenging task that can be accelerated by imitation learning. Recent approaches based on neural network policies were able to present promising results in generating human-like driving behavior. However, one drawback is that certain behaviors, such as avoiding accidents, cannot be guaranteed with such policies. Therefore, we propose to combine recent imitation learning methods like GAIL with a rule-based safety framework to avoid collisions during training and testing. Our method is evaluated on highway driving scenes where all vehicles are controlled by our driving policies trained on the real-world driving dataset highD. In this setup, our method is compared to a standard neural network policy trained with GAIL. Agents using our method were able to match GAIL performance while additionally guaranteeing collision-free driving.