An adaptive multi-head attention framework for unmanned surface vehicles formation control and obstacle avoidance

Formation control of multiple unmanned surface vehicles (USVs) is key to efficient marine exploration, environmental monitoring, and maritime rescue. However, dynamic formation size and complex sea conditions challenge the adaptability of formation control systems. To address this challenge, a multi-head self-attention multi-agent proximal policy optimization algorithm (MSA-MAPPO) is proposed for adaptive formation control of USVs in a leader-follower setting. The algorithm designs a multi-head self-attention structure to uniformly encode a variable number of follower USVs, thereby adapting to dynamic changes in formation size. A difference-reward mechanism is further used for credit assignment by decomposing the global formation objective into individual contribution signals, which promotes cooperative decision-making among USVs. In addition, multiple simulation environments are run in parallel to accelerate data collection, and curriculum learning is used to progressively increase the difficulty of formation-control tasks. Simulation results show that the proposed method is able to maintain a stable leader-follower structure and achieve safe obstacle avoidance under multiple obstacles, formation-size switching, and ocean-current disturbances, which demonstrates its effectiveness for adaptive formation and cooperative collision avoidance of USVs.