Hybrid Physics-Learning Model Based Predictive Control for Trajectory Tracking of Unmanned Surface Vehicles

The trajectory tracking control of unmanned surface vehicles (USVs) generally faces challenges from complex uncertain hydrodynamics and system constraints. In this paper, we leverage the capability of the Deep Neural Networks (DNN) in approximating arbitrary nonlinear dynamics, and propose a hybrid physics-learning model based predictive control (PL-MPC) method, Neural-sailing, for USVs. The first feature of PL-MPC is robust in that it handles the modeling errors caused by either the USV uncertain hydrodynamics or the environmental disturbances. Particularly, the hybrid physics-learning model combines a conventional mass-Coriolis-damping USV maneuvering model and a DNN model. The DNN model is trained with comprehensive USV motion data. The hybrid model enables efficient model-based control design with easily pre-obtained simple models. Moreover, the stability of the PL-MPC is guaranteed following the quasi-infinite nonlinear predictive control scheme. The second feature of PL-MPC is being computationally efficient. We propose a successive linearization approach to deal with the complexity of the prediction model caused by the introduction of DNN. The optimal control sequence computed from the previous step is utilized to obtain a linearization trajectory in the current step. This is the first time that a hybrid physics-learning model is successively linearized and used for predictive control. Simulation results reveal that PL-MPC has higher tracking precision than the conventional predictive control. The PL-MPC with successive linearizations has comparable tracking control performance with that of PL-MPC. However, with successive linearizations, the controller is much more computationally efficient, and achieves the trade-off between the control performance and computational burden.