Path Tracking Control for Autonomous Vehicles with GP-MPC Considering System Uncertainty

Dealing with system uncertainty is a significant challenge in model-based motion control for autonomous driving. To enhance the control performance of unmanned vehicles under uncertain operating conditions, a learning-based model predictive control (MPC) method is proposed in this paper. The uncertain of the prediction model is estimated by utilizing Gaussian Process Regression (GPR), enabling the controller to predict the vehicle’s future state even with uncertain parameters, thereby achieving optimal control. First, the vehicle’s residual model is learned using GPR to compensate for the unmodeled dynamics between the dynamic model and the actual vehicle. Next, the covariance matrix is introduced to quantitatively represent the uncertainty in the model’s predictions, and the matrix eigenvalues are used to tighten road boundary constraints, further improving control safety. Finally, a dSPACE hardware-in-the-loop experimental platform is established, and a comparative experiment between the proposed GP-MPC and a traditional nonlinear MPC (NMPC) is conducted in a typical continuous cornering scenario. The results demonstrate that the GP-MPC method outperforms the traditional NMPC in uncertain system model parameter conditions.