hGCBF: Heterogeneous Graph Control Barrier Functions for Safe Multi-Agent Systems

Provable safety remains a critical challenge for heterogeneous multi-agent systems (MASs), where differences in dynamics, state/action spaces, and sensing capabilities compromise existing learning-based methods that assume system homogeneity. To address this gap, we propose the heterogeneous Graph Control Barrier Function (hGCBF), the first neural control barrier function framework designed for heterogeneous MASs. Our approach introduces a novel extended system formulation via zero-padding to unify heterogeneous state-action spaces, upon which we formally define hGCBF and establish theoretical safety guarantees. We further design a heterogeneous graph neural network architecture with agent-type embeddings and a decoupled node/edge attention mechanism to jointly learn the hGCBF and a distributed controller, enhanced by type-specific loss weighting. Extensive experiments in large-scale 2D and 3D environments demonstrate that our approach achieves over 27% higher safety rates than homogeneous baselines at scale while maintaining strong performance under compositional shifts in agent-type ratios.