Coordinated Diffusion: Generating Multi-Agent Behavior Without Multi-Agent Demonstrations

Lasse Peters, Laura Ferranti, Andrea Bajcsy, Javier Alonso-Mora

[arXiv] [PDF]

@article{peters2026codi,
    title   = {Coordinated Diffusion: Generating Multi-Agent Behavior Without Multi-Agent Demonstrations},
    author  = {Peters, Lasse and Ferranti, Laura and Bajcsy, Andrea and Alonso-Mora, Javier},
    journal = {arXiv preprint arXiv:2605.11485},
    year    = {2026},
    url     = {https://arxiv.org/abs/2605.11485}
}

News

Abstract

Imitation learning is a powerful tool for teaching robots complex behavior, but collecting coordinated, multi-agent demonstrations is costly and often impractical. We present Coordinated Diffusion (CoDi), which couples independently trained single-agent diffusion policies through a user-defined multi-agent cost function, without requiring any coordinated demonstrations. We introduce a diffusion-based sampling scheme that combines single-agent policies with a cost-driven guidance term, supporting gradient-free estimation and hence non-differentiable cost functions. Across two-arm manipulation tasks, we show that CoDi learns coordinated behavior from single-agent data more efficiently than multi-agent baselines.

Code

Code coming soon.