Data-driven Vehicle Torque Vectoring Control Using Streaming Gaussian Process MPC

Abstract

This paper presents a torque vectoring control system that utilizes model predictive control (MPC) augmented by a Gaussian process (GP). Conventional MPC can suffer performance degradation when faced with un-modeled system dynamics or changing operating conditions. To address these limitations, the current work employs the GP to learn and compensate for residual vehicle dynamics and disturbances. The proposed framework features a dynamic online adaptation of the GP model, where its predictions are continuously refined based on recent driving data through a data buffering strategy and periodic hyperparameter re-optimization. This online learning framework, termed Streaming GP in this work, enhances overall system control accuracy and adaptability. The effectiveness of the proposed algorithm is demonstrated through comprehensive simulations on a vehicle model across various challenging driving scenarios, showing torque vectoring performance compared to conventional methods.