Spatially valid underwater target capture by multiple AUVs with dual policy reinforcement learning framework of target allocation and pursuit control

This paper proposes a dual-policy reinforcement learning framework for target allocation and pursuit control in spatially valid underwater target capture, enabling multiple autonomous underwater vehicles (AUVs) to form feasible subgroups and capture maneuvering targets in three-dimensional underwater environments. The allocation policy estimates AUV–target preferences and transforms them into feasible assignment results through a differentiable Sinkhorn projection with capacity constraints. The control policy network receives the assignment results and generates control commands for each AUV to achieve effective target encirclement. Subsequently, an allocation-aware relational encoder is designed as a shared GAT-based representation module for both policy networks, providing structured relational features by encoding interaction patterns induced by the current allocation. Ultimately, a spatially valid capture criterion is formulated by integrating geometric enclosure, formation quality, and dynamic closing advantage for valid underwater encirclement. Simulation results in three-dimensional underwater scenarios show that the proposed framework achieves higher capture success rates and better task efficiency than representative heuristic assignment and conventional reinforcement learning framework baselines. Scalability tests in a larger target scenario further demonstrate the stability and adaptability of the proposed method.