Reproduce “Exploring Zero-Shot Emergent Communication in Embodied Multi-Agent Populations” 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

Kalesha Bullard, Franziska Meier, Douwe Kiela, Joelle Pineau, Jakob Foerster

Embodied agents can learn emergent communication protocols that generalize zero-shot to unseen partners by leveraging universal physical constraints such as energy expenditure and non-uniform intent distributions. This approach enabled near-perfect zero-shot coordination for two intents when an observer was explicitly provided with latent energy features, moving beyond symbolic "cheap-talk" channels.

Key insights

  • Universal physical constraints, specifically energy cost and non-uniform intent distributions, can serve as grounding for emergent communication protocols that facilitate zero-shot generalization.
  • The discovery or explicit provision of this induced latent structure (e.g., energy values) to the observer is critical for achieving robust zero-shot coordination, especially as task complexity increases.
  • Optimizing for continuous-action, generalizable communication protocols in high-dimensional spaces is challenging due to non-convexity, sender-receiver lock-in, and the combinatorial problem of ordering continuous features.

Summary by alphaXiv.

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