Active disturbance rejection path tracking control of vehicles with adaptive observer bandwidth based on Q-learning

In this paper, a novel algorithm for vehicle path tracking control is introduced, focusing on maintaining tracking accuracy and minimizing steering wheel oscillation to enhance mechanism lifespan and passenger comfort. Vehicle kinematics model is utilized to formulate a second-order dynamic equation for lateral error, integrating yaw error into the standard first-order dynamic equation. A Proportional-Derivative (PD) controller is designed, incorporating an ‘extended state’ to compensate for the discrepancy between the model and actual vehicle dynamics, termed as the ‘total disturbance’. This ‘total disturbance’ is observed by an Extended State Observer (ESO), and a disturbance rejection law, combined with the PD controller, is employed to achieve the desired yaw rate. For improved vehicle safety and comfort, a dynamic constraint on the yaw rate, based on the vehicle’s motion and dynamic principles, is proposed. The vehicle’s nonlinear dynamics are addressed through feedback linearization, converting the target yaw rate into the required steering angle, which is then executed by the steer-by-wire system. An adaptive online algorithm for adjusting the ESO bandwidth, using Q-learning, is implemented. This optimization aims to balance tracking accuracy and steering wheel oscillation. A mathematical analysis confirms the stability of the time-varying bandwidth ESO and the overall system, ensuring limited estimation and control errors. Experimental comparison with the classical Stanley and Model Predictive Control (MPC) method demonstrates the algorithm’s effectiveness, maintaining lateral error within ±0.1 m.