Reproduce “Coordinated Diffusion: Generating Multi-Agent Behavior Without Multi-Agent Demonstrations” 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

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

Coordinated Diffusion (CoDi) synthesizes coordinated multi-agent behavior from readily available single-agent demonstrations, by composing independently pre-trained single-agent diffusion policies with a user-defined multi-agent cost function. The framework demonstrated robust collaboration for a two-arm manipulation task on hardware and outperformed multi-agent baselines in data efficiency and task performance in simulation.

Key insights

  • Complex multi-agent tasks often rely on low-level capabilities present in single-agent tasks, with coordination being the primary challenge, making single-agent data a valuable prior.
  • Coordinated multi-agent behavior can be effectively synthesized by leveraging single-agent policies for behavioral support and a multi-agent cost function for directing coordination.
  • The proposed gradient-free guidance score estimator, based on Monte Carlo integration from a Gaussian posterior approximation, is critical for effective coordination when guiding single-agent base policies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.