Dual Deep Neural Networks for Improving Trajectory Tracking Control of Unmanned Surface Vehicle

Abstract

This study proposes a deep learning-based method for trajectory tracking control of Unmanned Surface Vehicle (USV). Trajectory tracking control is an effective approach for autonomous sailing, making the USV to track towards a desired route. Dual Deep Neural Networks (Dual-DNN) are presented in this study to evaluate and revise the traditional Line-of-Sight (LOS) guidance algorithm. In particular, One DNN is used to evaluate the sailing effect of USV, while the other DNN is used to estimate the cross-track distance and lateral distance of the guidance law. The experimental results demonstrate that by adopting the proposed Dual-DNN model, the trajectory tracking error is reduced by 5.3% and 21.7% compared to the Single-DNN model and the traditional LOS model, respectively. The magnitude and frequency of throttle and rudder manipulations have been reduced. The smooth curves from the actuators are more consistent with the regular mode of surface vehicle maneuvering.