Hierarchical Deep Reinforcement Learning-Based Adaptive Task Allocation for Multi-AUV Cooperative Hunting

Cooperative hunting has emerged as a challenging task in complex underwater environments, involving multiple autonomous underwater vehicles (AUVs) jointly performing the subtasks of searching, tracking and capturing. However, existing approaches typically deal with these subtasks in isolation, lacking a unified and adaptive task allocation framework capable of responding to rapidly changing target behavior and limited communication. In this paper, we propose a novel hierarchical deep reinforcement learning (DRL) framework that jointly optimizes task allocation and action execution in an end-to-end manner. The high-level policy employs QPLEX to adaptively allocate each AUV to a specific task based on local observations and shared state information. The low-level policy, built on MADDPG, executes task-specific actions to perform the allocated tasks. By integrating discrete task allocation with continuous action control, the proposed framework effectively reduces action space dimensionality and enhances decision-making flexibility. Simulation results show that our method can effectively accomplish the multi-AUV cooperative hunting task, outperforming baselines in terms of cooperative performance and hunting efficiency.