Learning-based Nonlinear Model Predictive Control for a High Performance Virtual Driver

Recently, Model Predictive Control (MPC) has seen great progress in both research and many application areas, including autonomous driving vehicles. For autonomous race cars, Nonlinear MPC (NMPC) can be of great advantage, as it enables optimal control even at the physical limits of cornering speed and acceleration. However, to ensure reliable high-performance control, an accurate internal model is necessary. To this end, this manuscript describes a Learning-based NMPC approach that implements a grey-box model where the data-driven component is based on Gaussian Processes. Challenges arising from a space-discrete model formulation are addressed, and specific methods to reduce the computational burden, such as a sparse Gaussian Process approximation and a features selection procedure, are presented. The proposed method has been tested in a high-fidelity simulation environment, and it clearly outperforms the nominal model NMPC, both in open-loop predictions and in closed-loop performance.