Kyry10/h2-dances

H2 Dances

Unitree H2 dance dataset: 1363 dances (5.0 hours) from the AIST Dance Video Database, retargeted to the 31-DoF H2.

Website · 3D viewer · 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	1363
Styles	10
Total length	299.5 minutes at 60 fps
Robot	Unitree H2, 31 DoF
Dance source	AIST Dance Video Database / AIST++ (arXiv:2101.08779)
Retargeting	UMR (arXiv:2609.02134), H2 config
Tracking check	none: no public whole-body tracking controller for the H2 was available, so these are kinematic references only
Coverage	1363 of the 1363 usable AIST++ clips

Styles

House · 140	Break · 133	Waack · 125
Krump · 139	Pop · 140	Lock · 141
LA-style Hip-hop · 141	Middle Hip-hop · 141	Street Jazz · 136

Plus Ballet Jazz · 127. Numbers are dances per style. More videos on the website.

Use it for

• Motion-imitation RL on the Unitree H2: whole-body tracking targets for BeyondMimic-style trackers, mjlab, Isaac Lab or MuJoCo (convert qpos to your trainer's motion format).
• Training a general motion tracker for the H2: a large, varied set of fast, expressive reference motions.
• 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, or against the same dance on the Unitree G1 and Booster K1.

Quick start

import numpy as np
with np.load("motions/gBR_sBM_cAll_d04_mBR0_ch01.npz", allow_pickle=False) as m:
    qpos = m["qpos"]            # T x 38
    fps = float(m["fps"])       # 60
root_xyz  = qpos[:, :3]         # metres, world Z-up
root_wxyz = qpos[:, 3:7]        # unit quaternion, WXYZ
joints    = qpos[:, 7:]         # 31 joints, radians

Joint order is in joint_names.json and in each file's joint_names (Unitree H2 MuJoCo model from unitreerobotics/unitree_mujoco (BSD-3-Clause)). 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 38, float32), fps, joint_names, 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, quality metrics, SHA-256
joint_names.json	The 38 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 made

1. All 1363 usable AIST++ clips (the official list minus the AIST++ ignore list) were converted from SMPL to SMPL-X and retargeted to the H2 with UMR, at the original timing.
2. Every retargeted clip is included (1363). No tracking controller was run, so there is no guarantee that a given clip can be tracked by a physical or simulated H2.
3. Each clip was checked kinematically in MuJoCo; the ref_* columns of manifest.csv report self-collision, floor penetration, foot sliding, the fastest joint and joint-limit excursions. strict_quality_screen = True marks 397 clips that pass the same thresholds as G1 Dances and K1 Dances (quality_criteria.json).

Limitations

• Not checked with a controller. Unlike G1 Dances and K1 Dances, no clip here was run through a tracking policy, in simulation or on a robot. Expect some clips to be hard or impossible to track, especially fast floor work.
• Some clips contain self-collision, foot sliding or a fast joint flip (218 clips have a joint faster than 30 rad/s, see ref_max_joint_speed_radps). The checks are automated, not a frame-by-frame review.
• The robot model is Unitree H2 MuJoCo model from unitreerobotics/unitree_mujoco (BSD-3-Clause). Check the joint order and limits against your own model before use.

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++ and UMR (both entries are in CITATION.bib), and this dataset:

@misc{kolesnichenko2026h2dances,
  title={H2 Dances: AIST++ Dances Retargeted to the Unitree H2 Humanoid},
  author={Kolesnichenko, Kyrylo},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/Kyry10/h2-dances}}
}

Credits

Dances: AIST Dance Video Database (AIST) and AIST++ (Google). Retargeting: UMR. Robot model: Unitree H2 MuJoCo model from unitreerobotics/unitree_mujoco (BSD-3-Clause). Made by Kyrylo Kolesnichenko, ATR Lab, Kent State University. Same dances for other robots: G1 Dances (Unitree G1), K1 Dances (Booster K1).