Collision-avoidance under COLREGS for unmanned surface vehicles via deep reinforcement learning
Yong Ma, Yujiao Zhao, Yu‐Long Wang, Langxiong Gan, Yuanzhou Zheng
Abstract
Collision avoidance for unmanned surface vehicles (USVs) is significant for the fulfillment of autonomous navigation. Generally, classical collision-avoidance algorithms are proposed for relatively simple encounter situation, in this scenario only two USVs are stressed. Furthermore, to generate the rational manoeuvre operations, it is necessary that USVs should abide by International Regulations for Preventing Collision at Sea (COLREGS). However, COLREGS has not paid attention to rules for multi-USV collision-avoidance problem. Furthermore, those collision-avoidance rules in COLREGS have not been quantified for USVs. Following that, this paper utilizes deep reinforcement learning (DRL) algorithm to resolve collision-avoidance for USVs even in complex encounter situations. Within our DRL algorithm, related COLREGS is quantified properly and integrated into the DRL model, and then encounter situation of USVs is formulated as environmental observation value, accordingly a set of decision making is reached by decision-making neural network, and the reward function is designed for updating network parameters iteratively. Consequently, collision avoidance for USVs can be achieved eventually. By employing our DRL algorithm, collision avoidance for USVs under generous complex scenarios are resolved with the aid of corresponding intelligent decision-making operations. Simulation results verify the effectiveness of our DRL algorithm.