X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human Demonstrations

Maximus A. Pace* Prithwish Dan* Chuanruo Ning Atiksh Bhardwaj Audrey Du Edward W. Duan
Wei-Chiu Ma† Kushal Kedia†
Cornell University
X-Diffusion selectively incorporates noised human actions into Diffusion Policy training

Robots and Humans Behave Differently

Adding Noise Makes Human and Robot Actions Indistinguishable

X-Diffusion Pipeline

We convert human videos into robot-aligned state-action pairs using 3D hand-pose estimation and lightweight retargeting, and reduce the visual gap with object masks and keypoint overlays.

Adding noise suppresses embodiment-specific details while preserving task intent. We use a human-vs-robot classifier to find the earliest noise level where human actions are indistinguishable from robot actions and treat them as safe supervision.

Key Results

1) X-Diffusion can train on all human data, while naively training on all human data leads to infeasible robot motions

We compare X-Diffusion with FILTERED (robot-verified human demos only), NAIVE (all human data), and ROBOT (robot-only) and find consistent gains on every task.

2) X-Diffusion outperforms prior cross-embodiment learning baselines

We demonstrate empirically that X-Diffusion outperforms other cross-embodiment learning baselines across all tasks, which see either small performance gains or degraded performance when using all human data.
Naive vs X-Diffusion

Paper

Paper thumbnail

BibTex

@misc{pace2025xdiffusiontrainingdiffusionpolicies,
  title={X-Diffusion: Training Diffusion Policies on Cross-Embodiment Human Demonstrations}, 
  author={Maximus A. Pace and Prithwish Dan and Chuanruo Ning and Atiksh Bhardwaj and Audrey Du and Edward W. Duan and Wei-Chiu Ma and Kushal Kedia},
  year={2025},
  eprint={2511.04671},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2511.04671}, 
}