Lateral Motion Control for Obstacle Avoidance in Autonomous Driving Based on Deep Reinforcement Learning

To assist autonomous vehicles in confronting obstacle collision avoidance situations, a model-free lateral motion control method is proposed for vehicle collision avoidance based on deep reinforcement learning (DRL). Specifically, Twin Delayed Deep Deterministic Policy Gradients (TD3) is introduced as the main framework for training the lateral controller. A dynamic driving risk field is established to assess the driving risk by quantifying the impact of obstacles and roads on the host vehicle. Then, a reward function accounting for both driving risk and comfort is incorporated into the TD3 training framework to optimize steering performance. Finally, the TD3 agent learns to control steering maneuvers for obstacle avoidance in a fashion that maximizes cumulative rewards via trials and errors in the CARLA simulation environment. The training and testing results showed that the proposed controller could generate real-time, safe and comfortable steering maneuvers to avoid collision in dense traffic scenarios.