Data-Driven Adaptive Control of Unmanned Surface Vehicles Using Learning-Based Model Predictive Control

Sammanfattning: In this thesis, the subject of data-driven control of Unmanned Surface Vehicles (USVs) is explored. The control task is formulated through Nonlinear Model Predictive Path Following Control (NMPFC). System identification (SYSID) and Reinforcement Learning (RL) are employed to improve performance in a data-driven manner. The objectives were to assess the resulting controller’s path-following ability, as well as its adaptability to new environments, enabling the use of the controller on different USVs and under different conditions. The evaluationwas done in simulation and in experiments with the Saab Kockums AB’s Piraya vessel. Based on the results presented, NMPFC gives low-error solutions in both simulation and experiments but seems non-robust against disturbances and model mismatch. The simulation results of the learning-based methods showed that enabling SYSID on a USV with an incorrect initial model would identify the correct model in one update. Moreover, applying SYSID in experiments roughly halved the USV tracking error, compared to the usage of a model identified offline. Lastly, the RL implementation was found to increase performance in offline simulation, though less than SYSID. Moreover, the RL method computational times prohibited real-time control of the Piraya. This led to the method not being deployed in experiments.