Kyry10/g1-dances

G1 Dances

Unitree G1 dance dataset: 893 dances (2.8 hours) from the AIST Dance Video Database, retargeted to the 29-DoF G1. All 893 are trackable zero-shot by NVIDIA SONIC. Free, no login.

Website · Space · Required citation · AIST Dance Video Database

These dances belong to AIST. Every motion here comes from the AIST Dance Video Database, created by AIST (National Institute of Advanced Industrial Science and Technology, Japan), through the AIST++ 3D annotations by Google. We only retargeted the dancers' movements to a robot. If you use this data, you must cite Tsuchida et al., ISMIR 2019. Commercial use is subject to the AIST Dance DB Terms of Use.

At a glance

Dances	893
Styles	10
Total length	168.6 minutes at 60 fps
Robot	Unitree G1, 29 DoF
Dance source	AIST Dance Video Database / AIST++ (arXiv:2101.08779)
Retargeting	PhySINK from PHUMA (arXiv:2510.26236)
Trackable by SONIC	all 893, zero-shot: NVIDIA SONIC (arXiv:2511.07820), MuJoCo simulation, no fine-tuning

Styles

House · 107	Break · 38	Waack · 79
Krump · 96	Pop · 138	Lock · 100
LA-style Hip-hop · 124	Middle Hip-hop · 97	Street Jazz · 89

Plus Ballet Jazz · 25. Numbers are dances per style. More videos on the website.

Use it for

• Motion-imitation RL on the Unitree G1: whole-body tracking targets for BeyondMimic-style trackers, mjlab, Isaac Lab or MuJoCo (convert qpos to your trainer's motion format).
• Testing whole-body controllers such as SONIC on diverse, fast, expressive motion.
• Robot dance and animation: 10 styles from House and Break to Krump and Waack.
• Retargeting research: every clip keeps its AIST ID, so you can compare against the original AIST++ motion.

Quick start

import numpy as np
with np.load("motions/gPO_sBM_cAll_d10_mPO2_ch05.npz", allow_pickle=False) as m:
    qpos = m["qpos"]            # T x 36
    fps = float(m["fps"])       # 60
root_xyz  = qpos[:, :3]         # metres, world Z-up
root_wxyz = qpos[:, 3:7]        # unit quaternion, WXYZ
joints    = qpos[:, 7:]         # 29 joints, radians

Joint order is in joint_names.json (the NVIDIA SONIC G1 29-DoF MJCF order). Quick check: python load_motion.py motions/<id>.npz. These are motion references (targets for a tracking controller), not low-level robot commands.

What's inside

File	Contents
motions/<id>.npz	Retargeted reference: qpos (T x 36, float32), fps, plus retargeting metadata (time_map, infeasible_src, src_frame0_t, method, info)
previews/	Animated 3 s preview of every dance + metadata.csv; this is the table in Data Studio
manifest.csv	One row per dance: style, routine, dancer and music IDs, duration, SONIC tracking metrics, quality flags, SHA-256
joint_names.json	The 36 qpos columns and joint limits
summary.json	Counts and durations
quality_criteria.json	Definition of the stricter quality screen
CITATION.bib, LICENSE.md	Required citation and use conditions
assets/	Preview video and images for this page
SHA256SUMS	Integrity check: shasum -a 256 -c SHA256SUMS

IDs keep the AIST naming, e.g. gBR_sBM_cAll_d04_mBR0_ch01 = genre Break, basic routine, dancer 04, music mBR0, choreography 01.

How the dances were selected

1. 1,363 usable AIST++ clips were retargeted to the G1 with PhySINK (from PHUMA).
2. Each clip was run through NVIDIA's released SONIC G1 controller in full-clip MuJoCo simulation, no fine-tuning, 0 ms action delay. A clip passes if the robot completes it without falling (root height error ≤ 0.25 m, orientation error ≤ 1 rad) with mean local tracking error ≤ 50 mm.
3. All 893 passing clips are included. The sonic_* columns give per-clip tracking quality.
4. strict_quality_screen = True marks 123 clips that also pass a stricter check for self-collision, floor penetration and foot sliding (quality_criteria.json).

Limitations

• Simulation only, not tested on a real robot. With a 30 ms action delay, success on held-out dancers dropped from 64.7% to 17.3% in our tests, so real-robot results may be much lower.
• Some clips contain brief self-collision or foot sliding. The checks are automated, not a frame-by-frame review.

Required citation and use conditions

The dances are AIST's. This dataset is released under the conditions set by the AIST Dance DB team:

1. Everyone who uses or redistributes this data must cite Tsuchida et al., AIST Dance Video Database, ISMIR 2019, and keep this requirement when redistributing.
2. Commercial use is subject to the AIST Dance DB Terms of Use.
3. No original AIST content (video, music, AIST++ SMPL annotations) is included or licensed here.

@inproceedings{tsuchida2019aist,
  title={AIST Dance Video Database: Multi-genre, Multi-dancer, and Multi-camera Database for Dance Information Processing},
  author={Tsuchida, Shuhei and Fukayama, Satoru and Hamasaki, Masahiro and Goto, Masataka},
  booktitle={Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR)},
  year={2019}
}

Please also cite AIST++, PHUMA and SONIC (all entries are in CITATION.bib), and this dataset:

@misc{kolesnichenko2026g1dances,
  title={G1 Dances: AIST++ Dances Retargeted to the Unitree G1 Humanoid},
  author={Kolesnichenko, Kyrylo},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/Kyry10/g1-dances}}
}

Credits

Dances: AIST Dance Video Database (AIST) and AIST++ (Google). Retargeting: PHUMA / PhySINK. Tracking check: NVIDIA SONIC. Robot model: Unitree G1. Made by Kyrylo Kolesnichenko, ATR Lab, Kent State University. Also see SONIC-K1 for the Booster K1.