Collaborative Disaster Rescue Strategy for Multi-UAV Systems Based on GOAPF-MARL and Proactive Replanning

Abstract

To address multi-UAV task allocation and dynamic obstacle avoidance in post-disaster scenarios, this paper proposes a hierarchical GOAPF-MARL (Goal-Oriented Artificial Potential Field Guided Multi-Agent Reinforcement Learning) framework. In the perception layer, a task model is built to couple casualty level and disaster scale, forming a unified task weight W and enabling more accurate demand representation under varying conditions. For upper-layer allocation, an improved SADCK-Medoid method is used with the Balanced Residual Capacity (BRC) metric to enable nonuniform UAV deployment across five rescue regions. In the execution layer, a Goal-Oriented Artificial Potential Field (GOAPF) is embedded into multi-agent reinforcement learning (MARL), improving navigation around buildings, debris, and expanding fire zones while enhancing coordination and decision stability in complex scenes. Simulation results demonstrate improved efficiency and robustness compared with baseline methods under both ideal and challenging settings.