Learning-based Nonlinear Model Predictive Control with application to quadrotor control

This thesis aims to apply Learning-based Nonlinear Model Predictive Control (Lb-NMPC) to the case of quadrotor control. In particular, within this context, the focus is on “learning the system dynamics", i.e. on applying machine learning tools to derive or improve the system model on which the controller relies, allowing to enhance the performances in presence of parameter uncertainty and unmodeled dynamics. Both blackbox and graybox modeling approaches are addressed, and attention is given to tailoring these strategies to the specific considered case study, by taking advantage of considerations based on the physics of the system. The resulting learning problems are then solved by considering two techniques: Gaussian process regression and feedforward neural networks. Finally, the derived learning-based controllers are implemented in Matlab and Simulink leveraging a tool for Nonlinear Model Predictive Control developed at the University of Padova, MATMPC. Care is then put in analyzing the results by comparing the metrics characterizing the performance of all the proposed solutions both from the learning and the control point of view.