Multi-path Following for Underactuated USV Based on Deep Reinforcement Learning

Recently, deep reinforcement learning (DRL) has attracted much attention in learning control policies for unmanned surface vehicles (USV) path following. However, existing DRL methods can only learn to follow one target path, and generalize poorly to other paths. This paper proposes a novel DRL based multi-path following method. Based on the generalized hindsight mechanism, our method can effectively reuse data generated in one path for training on other paths. Experiments show that the control policy trained by our method generalizes well to a wide range of paths, with significantly higher accuracy and success rate.