Role-Aware Ego-Centric Graph Attention Networks for Heterogeneous Multi-Agent Quadrotor Navigation in Partially Observable 3D Environments

This paper introduces a framework for navigating autonomous quadrotors in partially observable 3D environments. It employs multi-agent reinforcement learning with ego-centric graph attention to help agents, such as hunters, guards, and communicators, coordinate effectively. For each agent, the model generates dynamic ego-centric graphs that consider their proximity and the importance of their roles. Using the MAPPO model with generalized advantage estimation, agents can focus on the most relevant information. In simulation tests, the framework outperformed the global graph-based MARL model in success rates, collisions avoided, role-based rewards, and computational efficiency. These results show that the proposed quadrotor framework can adapt successfully to complex, partially observable environments.