Reinforcement Learning-Based Path Tracking Control for Connected Vehicles Considering Obstacle Avoidance

This paper proposes a dual-layer control framework to address vehicle path tracking in fleet control, aiming to enhance control performance and reduce computational burden. The framework divides vehicle control into lateral and longitudinal components, with the upper layer employing a longitudinal Deep Deterministic Policy Gradient (DDPG) and a lateral Deep Q-Network (DQN) algorithm, while the lower layer consists of a fuzzy PID controller. Additionally, to address obstacle avoidance requirements in path tracking, a Constant Avoidance Angle (CAA) algorithm is designed and can switch to lateral control when necessary. The results demonstrate that the controller rapidly converges vehicle errors to zero and maintains continuous path tracking after obstacle avoidance.