GMoE-Net: Gated Mixture-of-Experts for Heterogeneous Multi-Agent Motion Prediction

Motion prediction in scenes with diverse, interacting agents is a fundamental challenge for autonomous systems. Current methods often employ a single, homogeneous model, overlooking the distinct behavioral patterns of different agent types and showing fragility to real-world data imperfections like missing observations. To address these issues, this study proposes GMoE-Net, a robust and adaptable graph neural network. The framework first enhances trajectory data integrity via a tracking and interpolation module. It then uses an incremental, agentcentric coordinate system to create a robust feature representation. At its core, a gated mixture-of-experts (GMoE) decoder utilizes a dynamic gating network to route agent-specific features to different “expert” predictors, each specialized in a particular behavioral mode. This allows GMoE-Net to learn tailored models for heterogeneous agents within a single, end-to-end architecture. Extensive simulations on Argoverse demonstrate that GMoE-Net significantly outperforms state-of-the-art methods, proving the value of its robust design and explicit modeling of agent heterogeneity.