Deep Reinforcement Learning-Based Pursuit–Evasion Strategy for USVs in Complex Multi-Obstacle Environments

With the widespread application of Unmanned Surface Vehicles (USVs) in marine monitoring and military reconnaissance, collaborative pursuit in complex and dynamic environments faces significant challenges. Traditional path planning methods (e.g., A* algorithm, artificial potential field) suffer from limited adaptability to dynamic obstacles and difficulty in multi-objective collaborative optimization, while existing deep reinforcement learning approaches often neglect partial observability and variable-dimensional observations, resulting in insufficient strategy generalization. To address these challenges, this paper proposes a POSG-based deep reinforcement learning framework for multi-USV pursuit-evasion under partial observability, in which an enhanced Multi-Agent Soft Actor-Critic (MASAC) algorithm serves as the core learning component. By constructing hybrid observation spaces and a hierarchical reward function, the framework guides pursuer USVs to efficiently approach the target and form stable encirclement formations, while forcing the evader to adjust its escape trajectory in real time to delay capture. Through dynamic feature-compression encoding and obstacle-interaction pattern modeling, the proposed method alleviates the generalization challenge of policy networks in partially observable scenarios with varying input dimensions. Experimental results show that the proposed framework reaches stable convergence after approximately 3000 training episodes and achieves an 85% capture success rate over 100 evaluation episodes in unseen multi-obstacle maps, indicating improved capture efficiency and strong generalization under partial observability.