A dual-USV cooperative collision avoidance method based on deep reinforcement learning with reward redistribution mechanism

This is the authors' abstract. We don't add key points for this paper.

This paper proposes a deep reinforcement learning approach for cooperative collision avoidance in dual USVs, ensuring safe, smooth and rapid formation movement while maximising the joint sensing area. Independent policy models are developed for each USV, enabling them to generate collision avoidance actions and provide reference actions for their partner. A formation reward, combining obstacle avoidance and collaboration rewards, is designed to balance collision avoidance and collaborative sensing. To ensure fair reward allocation, an innovative redistribution mechanism evaluates performance differences between the two models, then redistributes the formation reward based on these differences, preventing suboptimal solutions and non-convergence during training. Experiments on the ROS + Gazebo simulation platform demonstrate that the proposed method enhances safety, collision avoidance speed, navigation smoothness and coverage of collaborative sensing areas compared to other methods.