École Nationale Supérieure de
Techniques Avancées

Résumé

In recent years, path planning using safe reinforcement learning has gained significant attention as an effective approach to ensuring that agents can navigate complex environments while adhering to safety constraints. Safe reinforcement learning combines the exploration capabilities of reinforcement learning with the assurance of safety, enabling agents to learn optimal strategies for reaching a goal without compromising safety. This paper presents a novel contribution to this field by developing a safety filter for reinforcement learning (RL) algorithms, utilizing a game-theoretic Model Predictive Control (MPC) approach. The proposed method addresses the challenge of ensuring the robustness of RL agents operating in semi-dynamic environments, where real-world conditions often deviate from those encountered during training. By integrating multi-step safety prediction with a multi-step search mechanism, this method surpasses traditional techniques such as action projection by providing improved system stability and minimizing the need for fallback safety policies. The comprehensive experiments conducted in this study demonstrate that the proposed method significantly enhances both performance and safety in unpredictable and constantly changing environments. Future work will focus on optimizing the balance between safety and performance, particularly in dynamic environments with obstacles, and reducing the likelihood of agents becoming trapped in local optima, thereby further enhancing the method's applicability to real-world scenarios.