Resilience-Driven Topology Reconfiguration via Hierarchical Deep Reinforcement Learning in Low-Altitude UAV Networks

In low-altitude wireless networks, Unmanned Aerial Vehicles (UAVs) are vulnerable to environmental disturbances, which can cause failures, disrupting network topology and weakening coverage and backhaul. Consequently, network reconfiguration has become an urgent problem. Such reconfiguration must jointly consider coverage and backhaul, while obstacles in low-altitude environments further increase its complexity and challenges. To address this problem, we propose UR-HDRL (UAV network Reconfiguration based on Hierarchical Deep Reinforcement Learning), a novel framework that adopts a hierarchical architecture to decouple safety constraints from communication performance optimization. The algorithm integrates Control Barrier Functions (CBFs) and Graph Neural Networks (GNNs) to ensure safety and enhance collaborative decision-making in environments with obstacles. Experimental results indicate that UR-HDRL achieves significant improvements in data transmission efficiency, network coverage, and collision avoidance compared with baseline methods. The results also reveal distinct differences between communication coverage and sensing coverage, highlighting the inherent trade-offs between them.