GE-HAPPO: Graph-Enhanced Heterogeneous-Agent PPO for Coverage Optimization in Communication-Constrained USV-UAV Teams

Conventional multi-agent reinforcement learning (MARL) methods for heterogeneous unmanned surface vessel-unmanned aerial vehicle (USV – UAV) coordination often rely on fixed-size observation vectors, which fail to capture dynamic spatial relationships, communication constraints, and heterogeneous capabilities. This limits their ability to achieve efficient coverage and safe operation in complex maritime-aerial environments. To address this, this paper proposes a Graph-Enhanced Heterogeneous Agent Proximal Policy Optimization (GE-HAPPO) framework that integrates a graph attention network into the MARL pipeline. The model constructs agent-centric heterogeneous graphs encoding communication, coverage, and spatial relations, and employs edge-aware multi-head attention to extract structural embeddings for policy optimization under a decentralized execution, centralized training scheme. In a constrained multi-target coverage scenario, GE-HAPPO improves coverage efficiency by 29.8%, task completion rate by 267%, and reduces island collisions by 71.4% compared to the baseline framework, demonstrating its effectiveness in structurally complex, constraint-rich multi-agent coordination tasks (This work was supported by the National Natural Science Foundation of China under Grant 62303297, the Shanghai Sailing Program under Grant 23YF1413100, the Open Research Project of the State Key Laboratory of Industrial Control Technology, China, under Grant ICT2024B40, and Shanghai Technical Service Center of Science and Engineering Computing, Shanghai University.).