Autonomous Docking of the USV using Deep Reinforcement Learning Combine with Observation Enhanced

Abstract

Unmanned Surface Vessels (USVs) have great potential for military and civilian applications as an unmanned device at sea that can perform complex and dangerous tasks in place of humans. This paper is concerned with the problem of autonomous docking for USVs under a dynamic environment. An End-to-End deep reinforcement learning autonomous docking (EDRLAD) algorithm is proposed to overcome the problems caused by the complicated control law in the traditional analytic approach without any mediated perception. The proposed observation enhanced lead together to faster convergence and more robust autonomous docking, speed correction, and attitude adjustment which using RGB image from a forward USV image sensor and USV's status sensor as input. The proposed algorithm based on reinforcement learning shows significant advantages in terms of utility and scalability compared to analytical methods. Using the same EDRLAD algorithm, targets and constraints can be highly customized to meet a variety of special requirements. Extensive experiments have been carried out to prove the effectiveness and simplicity of the algorithm.