Learning-based Model Predictive Control for Automotive Applications

In recent decades, the automotive industry has largely adopted advanced controllers, such as advanced driver assistance systems (ADASs), traction control systems, and anti-lock braking systems. Moreover, the significance of virtual prototyping tools has grown substantially in vehicle and controller design. These applications often depend on a controller responsible for generating vehicle input signals, e.g., accelerator, brake, and steering wheel angle. Within this context, Nonlinear Model Predictive Control (NMPC) has been widely employed due to its systematic approach in handling constrained multi-input multi-output systems. However, implementing NMPC strategies involves the necessity for a vehicle model, including complex subsystems like tires and suspensions. Hence, the model must be tailored to suit real-time applications while ensuring enough precision for achieving high-performance control. Furthermore, the versatility of NMPC controllers introduces a large number of parameters that require careful tuning, posing a significant challenge in NMPC design. To deal with these issues, Learning-based Nonlinear Model Predictive Control (LbNMPC) could be exploited. It is an innovative control strategy that integrates NMPC techniques with data-driven methods, allowing to leverage the advantages of both the methodologies. In its framework, different categories have been explored. On one side, the learning dynamics branch aims at obtaining or refining the dynamics model by using data-driven techniques. On the other side, the learning design branch exploits learning-based strategies to maximize the closed-loop performance of the underlying NMPC controller, by algorithmically tuning the NMPC parameters. In this thesis, implementations of LbNMPC methods for two-wheels and four-wheels vehicles are presented. On the learning dynamics branch, the development of grey-box and black-box dynamics models for both two-wheels and four-wheels vehicles is detailed, with specific focus on meeting the real-time constraint. Gaussian Process (GP) Regression has been chosen to model the data-driven components of the dynamics, due to the possibility of limit the dimensionality while preserving the model characteristics. Different offline and online reduction techniques have been explored, e.g., feature selection procedures, sparse GP approximations, and local GP approximations, and they have been adapted to the specific case of interest. On the learning design branch, the application of a genetic algorithm to obtain a high-performance virtual rider that considers the sensitivity with respect to parameters changes is described. The results show proficiency of the presented methodologies in both realizing data-driven model dynamics for control purposes while maintaining real-time applicability, and in obtaining an appropriate tuning of the NMPC controller for a virtual motorcycle. Within the former strategies, the development of grey-box models resulted very effective in improving the tracking performance of an NMPC controller for a virtual motorcycle, and in enhancing the efficacy of a nonlinear model predictive contouring controller for lap-time minimization for a virtual high-performance car. Moreover, the implementation of a LbNMPC based on a black-box model for a real go-kart shows the applicability of the strategy using purely data-driven models in a challenging scenario. Finally, the data-driven tuning of the virtual rider allowed to establish the trade-off between performance and robustness of the controller.