Semantic-Triggered Graph Reinforcement Learning for Multi-UAV Cooperative Coverage Control

Abstract

This paper addresses the challenge of cooperative coverage control for multiple Unmanned Aerial Vehicles (UAVs) operating under communication constraints and partial observability. Existing graph-based reinforcement learning methods face two primary challenges: limited perception caused by local observations and excessive reliance on inter-UAV communication resources. To address these issues, this paper proposes a semantic-triggered graph reinforcement learning framework for communicationefficient multi-UAV cooperative coverage. The framework employs a Semantic-Triggered Deep Recurrent Graph Network (ST-DRGN) to extract robust state representations from local observations, neighboring UAV interactions, and historical features. These representations then enable each UAV to learn a shared policy that aggregates task-relevant neighbor information through graph attention and activates inter-UAV communication according to the observed points of interest (PoIs). A communication-aware reward aligns local decisions with the global coverage-fairnessenergy (CFE) objective. Simulation results demonstrate that ST-DRGN outperforms representative baselines, achieving a superior coverage-fairness-energy trade-off with reduced communication overhead.