Efficient Uncertainty Propagation in Model-Based Reinforcement Learning Unmanned Surface Vehicle Using Unscented Kalman Filter

This article tackles the computational burden of propagating uncertainties in the model predictive controller-based policy of the probabilistic model-based reinforcement learning (MBRL) system for an unmanned surface vehicles system (USV). We proposed filtered probabilistic model predictive control using the unscented Kalman filter (FPMPC-UKF) that introduces the unscented Kalman filter (UKF) for a more efficient uncertainty propagation in MBRL. A USV control system based on FPMPC-UKF is developed and evaluated by position-keeping and target-reaching tasks in a real USV data-driven simulation. The experimental results demonstrate a significant superiority of the proposed method in balancing the control performance and computational burdens under different levels of disturbances compared with the related works of USV, and therefore indicate its potential in more challenging USV scenarios with limited computational resources.