TAG-MAPPO: 3D Multi-UAV Pursuit-Evasion via Type-Aware Graph Attention Networks

Abstract

3D multi-UAV pursuit-evasion games are a fundamental problem in cooperative autonomous systems. However, coordinating heterogeneous swarms, such as long-range Scouts and short-range Interceptors, in obstacle-rich 3D spaces remains challenging due to asymmetric kinematic limits, heterogeneous sensing capabilities, and obstacle-induced sensing occlusions. Traditional Heterogeneous Graph Neural Networks (HGNNs) structurally isolate parameter spaces for different types of agents, which can limit parameter sharing and data efficiency in continuous physical coordination tasks. To address this issue, we propose a lightweight type-aware framework, Type-Aware Graph Attention Multi-Agent Proximal Policy Optimization (TAG-MAPPO). TAG-MAPPO combines a shared graph backbone for universal kinematic representation with an early-modulation layer placed before message passing, allowing agent-type information to affect subsequent graph attention weights. Simulation results show that TAG-MAPPO achieves competitive capture performance while reducing episode length, collision rate, and energy cost compared with representative baselines. We further evaluate zero-shot transfer across different pursuer population sizes, showing structural adaptability to UAV dropouts and reinforcements without retraining.