Multivariate Gaussian process-based learning model predictive control with unscented Kalman filter for autonomous...

Modeling the nonlinear dynamics of autonomous surface vehicles (ASVs) is a complex challenge, driven by the intricate interplay of hydrodynamic effects and environmental uncertainties. In response to this challenge, this paper develops the system state and observation dynamics for ASVs using multivariate Gaussian process regression (MVGPR), then designs a learning-based model predictive control (MPC) scheme for trajectory tracking of ASVs. First, we introduce the application of MVGPR to model the ASVs dynamics, enabling accurate multi-input and multi-output correlation and uncertainty estimation, addressing the limitations of traditional Gaussian process regression (GPR) in high-dimensional settings. Based on the learned models, an unscented Kalman filter (UKF) is designed to improve state estimation accuracy through prior prediction and posterior updating, ensuring robustness even under unmeasurable states. Additionally, considering the impact of denial-of-service (DoS) attacks in communication networks, an MVGPR-based learning MPC framework is developed. By leveraging predictive capabilities, this framework eliminates the need for external compensators. The proposed method achieves robust and precise trajectory tracking while improving system stability under complex and uncertain maritime environments. Finally, the effectiveness of the proposed learning-based MPC algorithm is verified through comparative simulations and hardware experiments.