Can you trust your autonomous car? Interpretable and verifiably safe reinforcement learning

Safe and efficient behavior are the key guiding principles for autonomous vehicles. Manually designed rule-based systems need to act very conservatively to ensure a safe operation. This limits their applicability to real-world systems. On the other hand, more advanced behaviors, i.e., policies, learned through means of reinforcement learning (RL) suffer from non-interpretability as they are usually expressed by deep neural networks that are hard to explain. Even worse, there are no formal safety guarantees for their operation. In this paper we introduce a novel pipeline that builds on recent advances in imitation learning and that can generate safe and efficient behavior policies. We combine a reinforcement learning step that solves for safe behavior through the introduction of safety distances with a subsequent innovative safe extraction of decision tree policies. The resulting decision tree is not only easy to interpret, it is also safer than the neural network policy trained for safety. Additionally, we formally prove the safety of trained RL agents for linearized system dynamics, showing that the learned and extracted policy successfully avoids all catastrophic events.