Learning-based Nonlinear Model Predictive Control for A Motorcycle Virtual Rider

Model predictive control (MPC) is an established technique for high-performance control that is used in many application fields, ranging from the automotive sector to the control of biological systems. In particular, the possibility of applying MPC in a real-time fashion has led to its successful implementation in demanding scenarios like car driving and motorcycle riding. However, the dissimilarity between the internal model and the real system often represents a crucial issue of MPC, especially when dealing with nonlinear systems. For that reason, new research in the field has been devoted towards the introduction of learning techniques into MPC frameworks, notably in order to reduce the mismatch between the model used by the controller and the controlled system itself. Such approaches belong to the general field of learning-based MPC (LbMPC). This thesis revolves around the previously cited themes of LbMPC. Specifically, learning techniques are explored in order to improve the nonlinear physics-based model of a motorcycle system to be used for Nonlinear Model Predictive Control. The objective is to design a Virtual Rider for high performance motorcycle riding. The main challenges are related to the determination of which parts of the dynamic model would benefit most from the addition of a learnt component and which learning strategy to pursue. Most of the effort has been devoted to Gaussian Process Regression strategies, with the exploration of both black-box and grey-box approaches and the investigation of sparse approximations in order to guarantee the real-time implementation of learning-based MPC.