Reproduce “PoCo: Policy Composition from and for Heterogeneous Robot Learning” 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

Lirui Wang, Jialiang Zhao, Yilun Du, Edward H. Adelson, Russ Tedrake

Researchers at MIT CSAIL introduce PoCo, a framework that leverages trajectory-level diffusion models to probabilistically compose robotic policies learned from heterogeneous data. This approach enables enhanced generalization for complex tool-use tasks, demonstrating up to 100% success in real-world scene generalization and outperforming individual or naive data-pooled policies.

Key insights

  • Trajectory-level diffusion models can represent robot policies as probability distributions, allowing for principled composition by summing their score functions at inference time.
  • Composing policies trained on different data sources (simulation, human demonstrations, real robot) or modalities effectively leverages their complementary strengths to enhance overall robustness and generalization.
  • Integrating learned policies with analytical cost functions (e.g., for smoothness or safety) at inference time enables fine-grained control and adaptation of robot behavior without retraining.

Summary by alphaXiv.

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