Adaptive Path Tracking Control Method for Autonomous Vehicles Based on Deep Reinforcement Learning

ABSTRACT Path tracking is crucial for autonomous vehicle stability, but traditional methods struggle under high‐speed and uncertain conditions, while deep learning approaches often lack generalization. To address these issues, this paper proposes an adaptive path trac‐king control method based on the twin delayed deep deterministic policy gradient algorithm. The method introduces a dynamic look‐ahead mechanism that adjusts the steering angle based on previewed curvature, effectively mitigating delayed responses in high‐curvature turns. Additionally, a piecewise reward function with safety‐based terminal penalties is designed based on the vehicle's internal states and relative pose with respect to the reference trajectory, enabling coordinated control of lateral dynamics, longitudinal motion and ride comfort. Experimental results show that at a target speed of 10 m/s, the method achieves a lateral tracking error below 0.1 m across multiple trajectories, representing an improvement of nearly 50% compared to a traditional model predictive control method. The velocity control error remains within 1 m/s, demonstrating coordination between lateral and longitudinal control and ensuring high tracking accuracy. Furthermore, even across a range of target speeds, the method maintains a lateral error within 0.15 m, exhibiting strong generalization and providing a solid foundation for real‐world deployment.