Safe reinforcement learning for high-speed autonomous racing

The conventional application of deep reinforcement learning (DRL) to autonomous racing requires the agent to crash during training, thus limiting training to simulation environments. Further, many DRL approaches still exhibit high crash rates after training, making them infeasible for real-world use. This paper addresses the problem of safely training DRL agents for autonomous racing. Firstly, we present a Viability Theory-based supervisor that ensures the vehicle does not crash and remains within the friction limit while maintaining recursive feasibility. Secondly, we use the supervisor to ensure the vehicle does not crash during the training of DRL agents for high-speed racing. The evaluation in the open-source F1Tenth simulator demonstrates that our safety system can ensure the safety of a worst-case scenario planner on four test maps up to speeds of 6 m/s. Training agents to race with the supervisor significantly improves sample efficiency, requiring only 10,000 steps. Our learning formulation leads to learning more conservative, safer policies with slower lap times and a higher success rate, resulting in our method being feasible for physical vehicle racing. Enabling DRL agents to learn to race without ever crashing is a step towards using DRL on physical vehicles.