Trajectory Tracking MPC with Online Model Adaptation Using Gaussian Process Regression and Clustering

This paper presents an Adaptive Model Predictive Controller (AMPC) for vehicle trajectory tracking. The proposed approach combines Model Predictive Control (MPC) with an online parameter learning algorithm based on a Gaussian Process Regression (GPR). The goal is to improve the tracking accuracy caused by model errors and condition changes. The vehicle's yaw intensification is considered as the adaptive parameter. Compared to prior research, clustered data dictionaries are introduced for the parameter, which are updated periodically according to the trajectory dynamics and the tracking performance. The data dictionaries allow the AMPC to memorize multiple driving conditions and to resort quickly to a previously learned parameter value if similar conditions are detected. The approach is evaluated in a simulation study based on CarMaker and Matlab/Simulink, where the vehicle is tested along several different closed-loop circuits with additional weight to emulate condition changes. These experiments demonstrate that the AMPC is able to adapt to repeated condition changes and to transfer learned behavior to unknown trajectories.