Deep Reinforcement Learning-Based MPC for Lateral Maneuver of Autonomous Vehicles

This paper presents an intelligent control framework for autonomous vehicles that integrates Model Predictive Control (MPC) with Deep Reinforcement Learning (DRL) to improve trajectory tracking performance under variable driving conditions. The proposed approach employs a Deep Deterministic Policy Gradient (DDPG) agent to dynamically adjust the weighting parameters of the MPC cost function in real time, enabling adaptive balance between tracking precision and steering smoothness. A NARX (Nonlinear AutoRegressive with eXogenous inputs) neural network model is used to capture the nonlinear and time-varying dynamics of the vehicle, providing accurate multi-step state predictions for the MPC. Simulation results demonstrate that the DRL-enhanced MPC significantly improves lateral tracking accuracy - up to 68% reduction in error - and enhances control stability compared with conventional fixed-weight MPC. The controller exhibits robustness to abrupt changes in road friction and external disturbances. Despite its simulation-based validation, the proposed framework highlights the feasibility of real-time adaptive control and provides a promising direction toward experimental implementation on embedded automotive platforms.