leeyngdo/AMASS_Retargeted_for_G1

AMASS Retargeted for Unitree G1 Humanoid

This dataset contains motions from AMASS retargeted to the Unitree G1 29-DOF humanoid using two complementary retargeting pipelines:

Variant	Pipeline	Source data	Frame rate	File format	Sequences
g1/NMR/	NMR — Neural Motion Retargeting	AMASS SMPL-X Neutral, Stage II	30 FPS	.npz	16,297
g1/GMR/	GMR — General Motion Retargeting (unitree_g1 29 DOF)	AMASS SMPL-X Gender-specific, Stage II	30 FPS	.pkl	17,029

Pick NMR for a smoother, learning-based retargeting; pick GMR for an IK-based retargeting tuned to G1 with extra body-link positions and a wider dataset coverage (incl. DanceDB and HUMAN4D).

• Target robot: Unitree G1 (29 DOF body + free-floating root).
• Source human motion: AMASS (Stage II only — *_stageii.npz). Subject shape fits (*_stagei.npz) are intentionally skipped.

Repository Layout

g1/
├── NMR/
│   ├── ACCAD/
│   ├── BMLmovi/
│   ├── ...
│   └── WEIZMANN/
└── GMR/
    ├── ACCAD/
    ├── BMLmovi/
    ├── ...
    └── WEIZMANN/

g1/NMR/<DATASET>/ contains one .npz per motion sequence (flat naming: <DATASET>__<subject>__<motion>_stageii.npz).

g1/GMR/<DATASET>/ mirrors the original AMASS directory tree, e.g. g1/GMR/CMU/15/15_02_stageii.pkl.

---

Variant 1 — NMR (g1/NMR/)

Neural motion retargeting with RayZhao/NMR, epoch_30.pth. Post-processed with a 4th-order Butterworth low-pass filter at 5 Hz (30 FPS), applied when sequence length > 15 frames.

File format (NMR)

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)
source_path	()	str	Absolute path of the source AMASS .npz at retargeting time

T is the number of frames at 30 FPS.

Loading (NMR)

import numpy as np
data = np.load("g1/NMR/CMU/CMU__15__15_02_stageii.npz")
dof   = data["dof"]                # (T, 29)
trans = data["root_trans"]         # (T, 3)
quat  = data["root_rot_quat"]      # (T, 4), w-first

Per-Dataset Counts (NMR)

Dataset	#sequences
ACCAD	252
BMLmovi	1863
BMLrub	3060
CMU	1981
CNRS	79
DFaust	129
EKUT	348
EyesJapanDataset	750
GRAB	675
HDM05	215
HumanEva	28
KIT	4231
MoSh	77
PosePrior	35
SFU	44
SOMA	69
SSM	30
TCDHands	62
TotalCapture	37
Transitions	110
WEIZMANN	2222
Total	16,297

Coverage Notes (NMR)

• 1 file unrecoverable: BMLmovi/Subject_49_F_MoSh/Subject_49_F_19_stageii.npz — the corresponding NPZ inside the official AMASS tarball is a truncated/corrupt archive (BadZipFile), so retargeting is impossible without a re-issued source.

---

Variant 2 — GMR (g1/GMR/)

IK-based retargeting with GMR (General Motion Retargeting) v0.2.0, target unitree_g1 (29 DOF), config smplx_to_g1.json. Mink/MuJoCo IK solver with automatic human-height scaling. Sequences are resampled to 30 FPS, the root XY origin is offset to the first frame, and the lowest body link is grounded at z=0.

File format (GMR)

Each .pkl is a Python pickle of a dict:

key	shape	dtype	description
fps	scalar	float	aligned frame rate (30.0)
root_pos	(T, 3)	float64	Root XYZ position in meters
root_rot	(T, 4)	float64	Root orientation quaternion (x, y, z, w) (scipy convention)
dof_pos	(T, 29)	float64	G1 joint angles in radians
local_body_pos	(T, 38, 3)	float32	Body-link XYZ positions in the root-local frame (from forward kinematics with identity root pose)
link_body_list	list[str]	—	Names of the 38 body links (matches local_body_pos axis 1)

T is the number of frames at 30 FPS. dof_pos ordering matches the G1 motor list printed by GMR (pelvis → legs → waist → arms).

Loading (GMR)

import pickle
with open("g1/GMR/CMU/15/15_02_stageii.pkl", "rb") as f:
    data = pickle.load(f)
dof_pos        = data["dof_pos"]         # (T, 29)
root_pos       = data["root_pos"]        # (T, 3)
root_rot       = data["root_rot"]        # (T, 4), x,y,z,w
local_body_pos = data["local_body_pos"]  # (T, 38, 3)
link_body_list = data["link_body_list"]  # list of 38 link names

Per-Dataset Counts (GMR)

Dataset	#sequences
ACCAD	247
BMLmovi	1864
BMLrub	3061
CMU	1983
CNRS	79
DanceDB	151
DFaust	129
EKUT	348
Eyes_Japan_Dataset	750
GRAB	1340
HDM05	215
HUMAN4D	148
HumanEva	28
KIT	4007
MoSh	77
PosePrior	35
SFU	44
SOMA	69
SSM	30
TCDHands	62
TotalCapture	37
Transitions	103
WEIZMANN	2222
Total	17,029

Coverage Notes (GMR)

• The GMR pipeline excludes a hand-curated set of motions known to be ill-posed for G1 retargeting. The default exclusion list filters out crawl / lying / stair sequences (crawl, _lie, upstairs, downstairs) and any motion path matching the GMR repository's assets/hard_motions/{0,1}.txt lists, which is why some Stage II files in the original AMASS archive are not present here.

---

Citation

If you use this data, please cite AMASS plus whichever retargeting pipeline you used:

@inproceedings{AMASS:2019,
  title={{AMASS}: Archive of Motion Capture as Surface Shapes},
  author={Mahmood, Naureen and Ghorbani, Nima and Troje, Nikolaus F. and Pons-Moll, Gerard and Black, Michael J.},
  booktitle = {International Conference on Computer Vision},
  year={2019}
}

@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}
}

@article{joao2025gmr,
  title={Retargeting Matters: General Motion Retargeting for Humanoid Motion Tracking},
  author={Joao Pedro Araujo and Yanjie Ze and Pei Xu and Jiajun Wu and C. Karen Liu},
  journal={arXiv preprint arXiv:2510.02252},
  year={2025}
}

@software{ze2025gmr,
  title={GMR: General Motion Retargeting},
  author={Yanjie Ze and João Pedro Araújo and Jiajun Wu and C. Karen Liu},
  year={2025},
  url={https://github.com/YanjieZe/GMR}
}

Plus the per-AMASS-subset citation for any subset you use — see each subset's original paper listed on the AMASS website.

License

Use of this data is governed by the original AMASS license — see https://amass.is.tue.mpg.de/license.html. Each AMASS sub-dataset retains its own license terms; you must comply with every sub-dataset license for any subset you download from here.

The retargeted motions are derivative of AMASS and inherit the same usage restrictions.