Research on Autonomous Vehicle Motion Control Based on Deep Reinforcement Learning and Neural Networks

Abstract

Traditional motion control for autonomous vehicles (AVs) predominantly relies on precise vehicle dynamics models. However, model mismatch often compromises control accuracy in highly dynamic and complex driving scenarios. This paper proposes an integrated control framework that synergizes data-driven methods with advanced motion planning. Specifically, for longitudinal control, the Deep Deterministic Policy Gradient (DDPG) algorithm is implemented to achieve adaptive and robust speed tracking. For lateral control, a Deep Neural Network (DNN) is trained using expert simulation data generated by a Stanley controller to handle non-linear path-following tasks. Furthermore, an enhanced Hybrid A* algorithm is introduced to perform both global and local path planning, ensuring kinematic feasibility. A comprehensive closed-loop control architecture is systematically established, bridging high-level decision-making with low-level dynamics constraints. Simulation results demonstrate the effectiveness of the proposed scheme in improving tracking precision and environmental adaptability.