Gaussian-Process Model Predictive Trajectory Tracking Control for Unmanned Surface Vehicles Under Model Uncertainty

Abstract

Model uncertainty and environmental disturbances can significantly degrade unmanned surface vehicle (USV) trajectory tracking when only a nominal kinematic model is used. This paper proposes a Gaussian-process model predictive control (GP-MPC) method for USVs under input-dependent disturbances. A three-degree-of-freedom discrete kinematic model is established, and Gaussian process regression is used to learn residual disturbances from sampled data. The GP posterior mean is embedded into the MPC prediction model as feedforward compensation. Simulations on straight-line tracking show that GP-MPC reduces the average position error from 0.124 m to 0.038 m and the maximum error from 0.287 m to 0.095 m compared with standard nonlinear MPC. The revised validation further includes longduration, circular, and S-shaped tracking cases, together with an analysis of training-sample density and distribution on GP uncertainty and generalization. The results demonstrate that datadriven disturbance learning improves prediction consistency and tracking robustness while preserving the constrained receding-horizon structure.