Multi-stage hierarchical multi-agent reinforcement learning for UAV-quadruped completing search and rescue

Search and rescue (SAR) operations in post-disaster environments often involve complex terrain, limited visibility, and high safety risks for human responders. While unmanned aerial vehicles (UAVs) are effective for rapid exploration and target localization, their limited payload capacity and endurance restrict their ability to perform sustained ground-level tasks. Quadruped robots, in contrast, are well suited for traversing unstructured terrain and carrying equipment, but they typically lack global situational awareness. Based on the multi-agent reinforcement learning (MARL) framework and hierarchical multi-stage reinforcement learning (HMRL) algorithm, this paper proposes a multi-stage hierarchical multi-agent reinforcement learning method (MHMARL). This method is introduced to train a heterogeneous multi-agent collaborative system consisting of a UAV and a robotic dog (ANYmal-C), enabling the UAV to guide the dog to complete SAR operations in post disaster rescue scenarios. The proposed approach is evaluated in a high-fidelity simulation environment built on Isaac Lab. Experimental results demonstrate that, compared with direct end-to-end MARL training, the proposed method achieves more stable training and reliable UAV-quadruped coordination across multiple SAR scenarios.