A Reinforcement Learning Module for Reducing Curve Trajectory Error in End-to-End Autonomous Driving

This paper proposes an approach that integrates a reinforcement learning module into a supervised learning-based end-to-end (E2E) autonomous driving system to address the trajectory error problem that occurs during curve driving in Target Point-based E2E models. The representative Target Point-based autonomous driving algorithm, Trajectory-guided Control Prediction (TCP), suffers from trajectory deviation issues in curved paths due to the shortcut problem. The proposed model minimizes trajectory errors in real time through a Soft Actor-Critic (SAC) reinforcement learning-based adaptive control correction module that utilizes the image features and control outputs of TCP, and employs a reward function specifically designed to enhance stability in curve driving scenarios. Experimental results using a proprietary driving dataset collected in the MORAI simulator demonstrate that the proposed model reduces trajectory errors by approximately 80% compared to the conventional TCP in curve situations and proves its robustness in various driving environments.