R3D: Revisiting 3D Policy Learning

Project Page | Paper | GitHub

This repository contains the pre-processed data for R3D, a framework for robust and scalable 3D imitation learning. R3D identifies the causes of training instabilities in 3D policy learning and proposes a new architecture coupling a scalable transformer-based 3D encoder with a diffusion decoder.

The dataset provides trajectories for challenging manipulation benchmarks, including:

• RoboTwin
• ManiSkill

Download

As specified in the official repository, you can download the pre-processed data using the Hugging Face CLI:

hf download eddie-cui/r3d --repo-type dataset --local-dir ./R3D/data

Citation

If you find this work or dataset useful, please cite:

@article{hong2024r3d,
  title={R3D: Revisiting 3D Policy Learning},
  author={Hong, Zhengdong and Wu, Shenrui and Cui, Haozhe and Zhao, Boyi and Ji, Ran and He, Yiyang and Zhang, Hangxing and Ke, Zundong and Wang, Jun and Zhang, Guofeng and Gu, Jiayuan},
  journal={arXiv preprint arXiv:2604.15281},
  year={2024}
}