HandX

This dataset release accompanies HandX: Scaling Bimanual Motion and Interaction Generation.

• Paper: https://arxiv.org/abs/2603.28766
• Project page: https://handx-project.github.io
• Code: https://github.com/handx-project/HandX

HandX provides high-fidelity bimanual hand motion data, MANO-format hand parameters, and fine-grained text annotations for dexterous hand motion generation research.

Release Contents

This release contains:

• HandX motion data in canonicalized joint-position format.
• MANO-format hand parameters.
• Source metadata describing the full merged dataset order.
• HandX-authored text annotations for ARCTIC and H2O.
• License and third-party notice files.

This release does not redistribute ARCTIC or H2O motion, pose, video, image, mesh, or annotation.

Files

train_can_pos_all_wotextfeat.npz
test_can_pos_all_wotextfeat.npz
train_mano.npz
test_mano.npz
handx_to_dataset_source_release.json
annotations/
  ARCTIC_text/
  H2O_text/
LICENSE
THIRD_PARTY_NOTICES.md
licenses/

The released .npz motion data contain 84,729 training entries and 9,648 test entries.

Data Format

Each .npz file is keyed by string indices such as "0", "1", and so on.

The canonicalized motion files contain dictionaries with:

motion: ndarray, shape (60, 2, 21, 3), float32
left_annotation: list[str]
right_annotation: list[str]
interaction_annotation: list[str]

The MANO files contain dictionaries with:

left_shape: ndarray, shape (60, 10), float32
right_shape: ndarray, shape (60, 10), float32
left_pose: ndarray, shape (60, 48), float32
right_pose: ndarray, shape (60, 48), float32
left_trans: ndarray, shape (60, 3), float32
right_trans: ndarray, shape (60, 3), float32

The ARCTIC/H2O annotation files contain HandX-authored text annotations with:

seq_name
frame_start
frame_end
split
annotations

Loading

The primary data files are NumPy archives. Download the repository snapshot and load the files with NumPy:

from pathlib import Path
import json
import numpy as np
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="alexzhang598/HandX",
    repo_type="dataset",
))

train_motion = np.load(root / "train_can_pos_all_wotextfeat.npz", allow_pickle=True)
train_mano = np.load(root / "train_mano.npz", allow_pickle=True)

motion_sample = train_motion["0"].item()
mano_sample = train_mano["0"].item()

with open(root / "handx_to_dataset_source_release.json", "r") as f:
    source_order = json.load(f)

print(motion_sample["motion"].shape)
print(mano_sample["left_pose"].shape)
print(source_order["train"]["0"])

License

HandX is released under the Snap Inc. Non-Commercial Research License in LICENSE.

This is a mixed-source dataset release. Some entries are derived from third-party datasets and remain subject to the applicable third-party licenses or terms. See THIRD_PARTY_NOTICES.md and the files under licenses/ before using or redistributing the data.

Important notes:

• GigaHands-derived entries are subject to CC BY-NC 4.0.
• HOT3D-derived hand data is governed by the HOT3D Dataset License Agreement and is subject to CC BY-NC-SA terms for hand annotations.
• HoloAssist-derived entries are subject to CDLA-Permissive-2.0.
• MANO model files are not redistributed here. MANO-derived parameters require compliance with the MANO license terms.
• ARCTIC and H2O source data are not redistributed here.

Citation

If you use HandX, please cite:

@inproceedings{zhang2026handx,
    title     = {HandX: Scaling Bimanual Motion and Interaction Generation},
    author    = {Zhang, Zimu and Zhang, Yucheng and Xu, Xiyan and Wang, Ziyin and Xu, Sirui and Zhou, Kai and Zhou, Bing and Guo, Chuan and Wang, Jian and Wang, Yu-Xiong and Gui, Liang-Yan},
    booktitle = {CVPR},
    year      = {2026},
}

Please also cite the applicable source datasets when using entries derived from them.