Enhancing Trajectory-guided Control Prediction of End-to-end Autonomous Driving by Reinforcement Learning

End-to-end autonomous driving models often exhibit limited scalability because they rely on expert demonstration data. Imitation learning techniques such as Trajectory-guided Control Prediction (TCP) face adaptability limitations in varied environments, necessitating extensive data collection and retraining for novel driving scenarios, including static obstacle avoidance. To address this issue, we propose a hybrid framework integrating a pre-trained TCP model with a reinforcement learning (RL) module. The RL agent leverages TCP's latent features and trajectory predictions as its state. By interacting with a simulated environment and optimizing a task-specific reward function, the agent acquires an obstacle avoidance policy, eliminating the need for additional expert data. Experiments in the MORAI simulator demonstrate that our approach reduces vehicle collisions by 53.9% and pedestrian collisions by 45.9% compared to the baseline TCP model, while achieving higher route completion. These findings suggest that augmenting TCP with RL enhances system robustness and enables progressive expansion of driving capabilities, particularly for obstacle avoidance, without additional expert data. While this work focuses on static obstacle avoidance, the proposed TCP with RL framework can be adapted to other driving scenarios by modifying the reward and simulation setup without collecting new expert demonstrations.