A Bi-GRU Network Based Collaborative Hunting Algorithm for Multi-Unmanned Surface Vehicle

In collaborative operations, unmanned surface ve-hicles (USVs) must strategize for multi-target missions, ne-cessitating the development of effective collaborative strategies for multi-USV systems. This paper integrates the concept of velocity obstacles with a distributed deep reinforcement learning approach based on multi-agent proximal policy optimization. The research commences with a limited set of observational features, translating the surhuntings of each USV into navigational and hunting feature observations. These observations are then directly mapped to the USV s' absolute actions via a deep reinforcement learning algorithm. Each intelligent USV within the system employs a bidirectional gated recurrent unit network to retain information about environmental and self-changes, effectively addressing the issue of dimensionality while selecting information crucial for strategy training, thereby enhancing the algorithm's training efficiency. Simulation indicates that our method significantly outperforms traditional reinforcement learning algorithms in terms of average steps, average speed, and average rewards. The algorithm enables inter-vessel collision avoidance during collaborative tasks and allows each intelligent USV to autonomously learn and accomplish hunting missions.