Reproduce “InterAgent: Physics-based Multi-agent Command Execution via Diffusion on Interaction Graphs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bin Li, Ruichi Zhang, Han Liang, Jingyan Zhang, Juze Zhang, Xin Chen, Lan Xu, Jingyi Yu and 1 more
InterAgent introduces the first end-to-end framework for text-driven, physics-based multi-agent humanoid control, leveraging a novel Interaction Diffusion Transformer and an interaction graph for fine-grained interaction modeling. The system generates highly coherent and physically plausible multi-agent behaviors, achieving superior text-motion alignment with an R-Precision of 0.375 (Top-1) and an FID of 0.582, outperforming existing kinematic and physics-based baselines.
Key insights
- Representing inter-agent dynamics with an Interaction Graph is more effective than simple relative states, providing richer structural information for coordinated actions.
- Sparsity is a valuable principle in modeling multi-agent interactions; selectively attending to critical joint-to-joint connections significantly improves robustness and efficiency.
- Decoupling different modalities (proprioception, exteroception, action) within a multi-stream transformer architecture mitigates cross-modal interference, leading to more coherent and physically plausible behaviors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.