Data-Driven Learning-Based Lane Keeping Robust Optimal Control of Autonomous Vehicles

Existing lane-keeping strategies either rely on precise system models to pursue optimal performance, or handle uncertainties at the expense of performance guarantees. Few methods successfully reconcile robustness with optimality, which is impractical for real-world implementation. To overcome these limitations, this paper develops a completely model-free H<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">∞</sup> control method based on reinforcement learning and adaptive dynamic programming (ADP) theory that operates without prior knowledge of system matrices. The proposed approach employs an online policy iteration algorithm to solve optimal control pairs while considering vehicle dynamics in both yaw and roll planes through a 4-degree-of-freedom (DoF) model incorporating lateral offset, heading error, and roll angle. A load transfer ratio (LTR) evaluates vehicle roll stability, and a generalized H<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">∞</sup> disturbance attenuation criterion is established using Kronecker products for online parameter learning. Simulation validation using CarSim-Simulink platform with a high-fidelity SUV model demonstrates that the algorithm achieves rapid convergence of lane keeping errors and roll angle to zero within 10 policy iterations. Compared to traditional sliding mode control, the proposed method exhibits superior trajectory tracking accuracy and faster response under challenging conditions including road slip, rollover prevention, and actuator saturation. The results provide a practical optimal control solution for autonomous vehicle lane keeping without requiring precise model information, significantly enhancing safety and reliability of autonomous driving systems.