Reproduce “Data Analogies Enable Efficient Cross-Embodiment Transfer” 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
Jonathan Yang, Chelsea Finn, Dorsa Sadigh
Research from Stanford University demonstrates that structured data analogies, particularly trajectory-paired demonstrations, are more effective than raw data volume or broad diversity for enabling cross-embodiment robot policy transfer. Policies trained with these analogies achieved up to 25% higher success rates in real-world tasks compared to those using large-scale unpaired datasets, also making previously untransferable skills viable for generalization.
Key insights
- Explicit data analogies, specifically trajectory-paired demonstrations, are essential for robust action-level transfer, particularly when dealing with end-effector morphology changes.
- Increasing data diversity broadly benefits perceptual generalization (viewpoint, appearance), but for morphology shifts, targeted data and strong pairing are indispensable.
- Combining reweighted, diverse open-source data with targeted trajectory-paired examples creates a compositional strategy that outperforms merely scaling unpaired data volume.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.