TGRL: A Trajectory-Guided Reinforcement Learning Approach for Robust Autonomous Racing

This paper presents a trajectory-guided reinforcement learning (TGRL) approach aimed at enhancing the robustness of autonomous racing by addressing model mismatch issues. Our proposed framework effectively mitigates these challenges by combining the heuristic strengths of reinforcement learning with the reliability of classical optimal planning methods. In this framework, the agent receives rewards based on the alignment of its actions with those selected by a classical planner that follows the optimal trajectory. Experimental results demonstrate that our method significantly improves generalization capabilities and robustness across various tracks, consistently outperforming baseline models under both normal and adverse conditions.