leeyngdo/Embody3D_Retargeted_for_G1
收藏Hugging Face2026-04-21 更新2026-04-26 收录
下载链接:
https://hf-mirror.com/datasets/leeyngdo/Embody3D_Retargeted_for_G1
下载链接
链接失效反馈官方服务:
资源简介:
---
license: other
license_name: embody3d
license_link: https://github.com/facebookresearch/embody3d
task_categories:
- robotics
tags:
- motion-retargeting
- humanoid
- embody3d
- smpl-x
- g1
pretty_name: Embody3D Retargeted for G1 Humanoid (NMR)
size_categories:
- 10K<n<100K
---
# Embody3D Retargeted for Unitree G1 Humanoid
This dataset contains motions from **Embody3D** retargeted to the **Unitree G1 29-DOF humanoid** using the **NMR (Neural Motion Retargeting)** pipeline from [Zhao et al. 2026](https://huggingface.co/RayZhao/NMR).
- **Source human motion**: Embody3D (per-frame SMPL-X joint positions + rotations, Z-up world) — 14 content categories.
- **Target robot**: Unitree G1 (29 DOF body + free-floating root).
- **Retargeting model**: NMR ([RayZhao/NMR](https://huggingface.co/RayZhao/NMR), `epoch_30.pth`).
- **Post-processing**: 4th-order Butterworth low-pass filter at 5 Hz (30 FPS), applied when sequence length > 15 frames.
## Repository Layout
```
NMR/
├── ACTING/
├── AIAGENT/
├── CHARADES/
├── CONTACT/
├── DAYLIFE/
├── EMO/
├── HAND/
├── ICEBREAKER/
├── INSTRUCT/
├── LOC/
├── LOCOMOTION/
├── POLYADICS/
├── PROXEMICS/
└── PT/
```
Each `NMR/<CATEGORY>/` folder contains one `.npz` per retargeted motion clip.
### File naming
The npz filename is the same as the source `.pkl` basename from the original Embody3D release, e.g.
```
c--20250108--1300--DXG448--SZM479--JON169--BWW760--pilot--MotionPrior--ACTING_Adult_Birthday_--103301-106600.npz
```
`<CATEGORY>` is derived from the content prefix of the source pkl filename (e.g. `ACTING_Adult_Birthday_` → `ACTING`). Mixed-case prefixes (e.g. `Hand_Contact`) are normalized to uppercase (`HAND`).
## Per-Category Counts
| Category | #clips |
|------------|-------:|
| ACTING | 1369 |
| AIAGENT | 872 |
| CHARADES | 2287 |
| CONTACT | 498 |
| DAYLIFE | 1301 |
| EMO | 2236 |
| HAND | 1077 |
| ICEBREAKER | 293 |
| INSTRUCT | 1835 |
| LOC | 2097 |
| LOCOMOTION | 2235 |
| POLYADICS | 1816 |
| PROXEMICS | 254 |
| PT | 882 |
| **Total** | **19,052** |
## File Format
Each `.npz` contains:
| key | shape | dtype | description |
|-----------------|-----------|---------|-------------|
| `dof` | `(T, 29)` | float32 | G1 joint angles in radians |
| `root_trans` | `(T, 3)` | float32 | Root XYZ position in meters (Y-up) |
| `root_rot_quat` | `(T, 4)` | float32 | Root orientation quaternion `(w, x, y, z)` |
`T` is the number of frames at **30 FPS**.
## Loading
```python
import numpy as np
data = np.load("NMR/ACTING/c--20250108--1300--DXG448--SZM479--JON169--BWW760--pilot--MotionPrior--ACTING_Adult_Birthday_--103301-106600.npz")
dof = data["dof"] # (T, 29)
trans = data["root_trans"] # (T, 3)
quat = data["root_rot_quat"] # (T, 4), w-first
```
## Conversion Notes
- The Embody3D pkl format ships **world-space joint positions and quaternions** per SMPL-X joint. NMR's preprocessor normally runs SMPL-X forward kinematics on `(transl, global_orient, body_pose)` to obtain the first 22 body joint positions; here we bypass that step and feed those positions directly from the Embody3D dict, using the pelvis quaternion as `global_orient`.
- Quaternion order in Embody3D is `(w, x, y, z)`, coordinates are **Z-up**. We convert to **Y-up** via `R = [[1,0,0], [0,0,1], [0,-1,0]]` to match NMR's internal convention, exactly as the AMASS loader does.
- **9 clips were skipped** because their length was below the 4-frame minimum the NMR chunked inference requires.
## Citation
If you use this data, please cite **both** Embody3D and NMR:
```bibtex
@inproceedings{embody3d,
title={Embody3D: A Large-scale Multimodal Motion and Behavior Dataset},
author={Meta FAIR and Codec Avatars Lab, Pittsburgh},
year={2025}
}
@article{zhao2026make,
title={Make Tracking Easy: Neural Motion Retargeting for Humanoid Whole-body Control},
author={Zhao, Qingrui and Yang, Kaiyue and Wang, Xiyu and Zhao, Shiqi and Lu, Yi and Zhang, Xinfang and Yin, Wei and Shen, Qiu and Long, Xiao-Xiao and Cao, Xun},
journal={arXiv preprint arXiv:2603.22201},
year={2026}
}
```
## License
Use of this data is governed by the original **Embody3D license**. The retargeted motions are derivative of Embody3D and inherit the same usage restrictions.
提供机构:
leeyngdo



