Reproduce “X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human 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
Maximus A. Pace, Prithwish Dan, Chuanruo Ning, Atiksh Bhardwaj, Audrey Du, Edward W. Duan, Wei-Chiu Ma, Kushal Kedia
Cornell University researchers developed X-DIFFUSION, a framework for training robot diffusion policies using cross-embodiment human demonstrations. The method selectively integrates human actions into training based on noise levels, allowing robots to learn manipulation skills despite dynamic incompatibilities. It achieved an average 16% improvement in task success rates over existing baselines across five real-world manipulation tasks.
Key insights
- Dynamically infeasible human actions can be conceptualized as 'noisy' versions of feasible robot actions, where embodiment-specific differences are masked by sufficient Gaussian noise.
- By selectively training on human demonstrations at high noise levels (beyond k*), a robot policy can extract valuable high-level task intent and object interaction cues without learning problematic low-level human execution details.
- This adaptive integration of human data allows for more comprehensive utilization of human demonstrations, even those deemed infeasible by manual filtering, transforming them into useful coarse guidance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.