893 dance motions for the Unitree G1 humanoid. All 893 are trackable zero-shot by NVIDIA SONIC in simulation. Here are 45 of them, five from each of nine styles.
Shown here: House, Break, Waack, Krump, Pop, Lock, LA-style hip-hop, Middle hip-hop and Street jazz. Ballet jazz is in the 3D viewer and the dataset.
Motions from the AIST Dance Video Database (AIST++), retargeted to the G1 with PHUMA and tracked by NVIDIA SONIC in simulation.
Watch 25 dance at once →See all 893 dances with videos on Hugging Face →
Free, no login. Each dance is one NPZ file: a 60 fps reference trajectory for the 29-DoF G1, ready as a target for motion-imitation RL or for testing a whole-body controller.
These are motion references, not low-level robot commands. Simulation only.
import numpy as np
from huggingface_hub import snapshot_download
path = snapshot_download("Kyry10/g1-dances", repo_type="dataset",
allow_patterns=["motions/*", "manifest.csv"])
with np.load(f"{path}/motions/gPO_sBM_cAll_d10_mPO2_ch05.npz") as m:
qpos = m["qpos"] # T x 36, 60 fps
root_xyz = qpos[:, :3] # metres, Z-up
root_wxyz = qpos[:, 3:7] # unit quaternion
joints = qpos[:, 7:] # 29 joints, radiansIf you use these motions, you must cite Tsuchida et al., AIST Dance Video Database, ISMIR 2019 (). Commercial use is subject to the AIST Dance DB Terms of Use. No AIST video, music or SMPL data is included. Simulation only, not tested on a real robot.