Learning-Augmented MPC for Autonomous Vehicle Path Tracking via Ensemble Residual Dynamics Learning

Accurate vehicle dynamics modeling is essential for path tracking control, especially under sharp-curvature or rapidly changing conditions where nonlinear and time-varying behaviors introduce significant discrepancies between the nominal model and real vehicle responses, ultimately degrading the performance of traditional Model Predictive Control (MPC). To address this challenge, this paper proposes a learning-augmented MPC framework that incorporates an ensemble learning-based Data-Driven Dynamics Refinement (DDR) Model to enhance predictive accuracy and control robustness. The DDR Model complements nominal vehicle dynamics by capturing complex behaviors that are difficult to represent analytically. An ensemble of independently trained neural predictors is employed to improve generalization performance and provide stable refinement across diverse driving conditions. Furthermore, a feature-driven activation mechanism is designed to selectively apply refinement only when pronounced nonlinear behaviors arise, thereby reducing unnecessary computational burden. High-fidelity simulation studies validate the effectiveness of the proposed method. In single- and double-lane-change scenarios, the refined dynamics reduce maximum lateral deviation by approximately 6 cm and 4 cm, and decrease the maximum vehicle heading error by 0.02 rad and 0.015 rad, respectively, demonstrating significant improvements in tracking accuracy and robustness.