Empowering Autonomous Underwater Vehicles Using Learning-Based Model Predictive Control With Dynamic Forgetting Gaussian Processes

Autonomous underwater vehicles (AUVs) present several challenges due to the complex and simultaneous interplay of various factors, including but not limited to unmodeled dynamics, highly nonlinear behavior, intercouplings, communication delays, and environmental disturbances. In particular, environmental disturbances degrade trajectory tracking performance for model-based controllers, e.g., model predictive control (MPC) algorithms. Data-driven methods such as the Gaussian process (GP) are effective at learning disturbances in real time; however, the underlying offline hyperparameter tuning process limits their overall effectiveness. To overcome this limitation, we propose a novel dynamic forgetting GP (DF-GP) methodology that compensates for operational disturbances, thus circumventing the need for hyperparameter retuning. In essence, the proposed method optimally combines the predictions of individual GPs—designed with handcrafted forgetting factors, rendering precise disturbance estimation of varying timescales. What is more, the predicted disturbances update the model parameters in MPC, facilitating a learning-based control framework that ensures accurate tracking performance in different underwater scenarios. Rigorous simulation and real-world experiments demonstrate the efficiency and efficacy of the proposed framework. The results show a 25% improvement in disturbance estimation and tracking performance, demonstrating that the proposed framework outperforms its direct competitors.