Manual approval required for data access. The complete Handwork-100 dataset has been uploaded and passed file-integrity verification. Access to dataset contents requires approval from OnePiece Labs. To request access, sign in to Hugging Face and submit an access request on this page. For commercial licensing, contact data@onepiecelabs.xyz.

Explore first: Browse the 30-minute Handwork-100-Preview, with approximately 10 minutes from each coordinate subset. Preview access is approved automatically after a signed-in request. It does not grant access to this full dataset.

Overview

Handwork-100 is a research dataset from OnePiece Labs for studying human manipulation, hand-object interaction and the spatial context of first-person demonstrations.

The collection contains 1,573 recording episodes and 7,068 PARTs, organized into three subsets by coordinate representation. It pairs stereo video with 3D hand annotations, calibration, timing information and task metadata. The world-coordinate subsets also provide head-mounted camera poses and SLAM-based trajectories.

Data is captured using a calibrated stereo head-mounted rig. Stereo videos retain their native 1600 × 1300 resolution and original frame timing. Duration is measured as unique recording time: left and right views are counted once.

Subsets and coverage

Subset	Pose representation	Hours	Episodes
`camera_frame_poses`	3D hand poses in camera coordinates	40.05	203
`world_frame_poses`	Hand and head-mounted camera poses in world coordinates	60.01	1,335
`gravity_aligned_world_poses`	Hand and head-mounted camera poses in gravity-aligned world coordinates	0.58	35
Total	Unique recording time	100.64	1,573

Durations are rounded for display; totals are calculated from unrounded values.

Camera-frame poses

Framewise left and right 3D hand annotations in the camera reference frame, paired with stereo video, calibration, timing information and task metadata. This subset supports analysis of hand articulation and manipulation relative to the wearer's view.

World-frame poses

World-coordinate hand annotations together with head-mounted camera poses and SLAM-based trajectories. This subset supports analysis of hand motion in relation to camera movement and the surrounding spatial reference.

Gravity-aligned world poses

Hand and head-mounted camera poses in gravity-aligned world coordinates. This subset supports analysis of manipulation and camera motion with an explicit vertical reference.

Scene coverage

The collection covers 20 scene categories. Demonstrations include sorting, organizing, packing, assembly, cleaning, tool use and other everyday and work-related activities.

Scene category	Hours	Share
Events and entertainment	10.31	10.24%
Home and domestic activities	9.47	9.41%
Arts and creative workshops	8.88	8.83%
Food and beverage	8.48	8.43%
Beauty and personal services	8.08	8.03%
Office and administrative services	7.44	7.39%
Miscellaneous	6.17	6.13%
Maintenance and repair services	4.84	4.81%
Education and professional training	4.48	4.45%
Sports and leisure activities	4.31	4.28%
Industrial production	4.10	4.08%
Emergency and public safety services	3.81	3.78%
Laboratory and scientific work	3.40	3.37%
Retail and consumer goods	3.03	3.01%
Printing and creative design	2.99	2.97%
General administration	2.80	2.79%
Apparel and fashion	2.48	2.46%
Agriculture and farm operations	2.28	2.26%
Construction and hardware services	2.19	2.17%
Energy and utility services	1.12	1.11%
Total	100.64	100.00%

Scene labels and recording-level task descriptions are retained in the release metadata. 02_DISTRIBUTION.json provides detailed scene and task statistics. Displayed values are rounded independently.

Data files and organization

MP4: original left and right stereo video.
HDF5: pose arrays and associated calibration, timing and validity fields.
JSON: episode metadata and recording identities.
Release indices: manifests, file lists, scene statistics and SHA-256 checksums.

Handwork-100/
├── README.md
├── 00_README.md
├── 01_FILE_LIST.tsv
├── 02_DISTRIBUTION.json
├── 03_CHECKSUMS.sha256
├── manifest.json
├── camera_frame_poses/
│   ├── manifest.json
│   ├── checksums.sha256
│   ├── episode_0/
│   └── ...
├── world_frame_poses/
│   ├── manifest.json
│   ├── checksums.sha256
│   ├── episode_0/
│   └── ...
└── gravity_aligned_world_poses/
    ├── manifest.json
    ├── checksums.sha256
    ├── episode_0/
    └── ...

Episode numbers start at zero within each subset. Use the subset name together with the episode number to identify an episode. The source_recording_id in episode metadata preserves the original recording identity.

Each episode contains its JSON metadata, HDF5 annotations and stereo video files. Multi-PART episodes retain their PART suffixes and file associations. Release manifests provide exact file paths and content checksums.

Download a small selection

After your access request is approved, authenticate with hf auth login. To begin with one episode rather than the full collection:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="onepiecelabs/Handwork-100",
    repo_type="dataset",
    local_dir="Handwork-100",
    allow_patterns=["README.md", "camera_frame_poses/episode_0/*"],
)

Replace the subset and episode path as needed. Removing allow_patterns downloads the full repository (approximately 790 GB).

LeRobot users: Download a prepared observation-only Preview group or follow the self-service conversion guide for the pinned native release. Tools and instructions are hosted with Preview; Preview access is automatic after a signed-in request, while this full collection requires separate manual approval. The validated reading profile uses LeRobot 0.6.1 with PyAV after download. No preconverted full 100-hour LeRobot copy is included.

Coordinate and annotation conventions

Hand pose describes hand joint annotations; head-mounted camera pose describes the position and orientation of the tracked camera reference. The subset name identifies the coordinate representation used for these annotations.

World frames are scoped to their native PARTs. Interpret each PART's trajectory using its supplied reference frame and calibration conventions. Gravity alignment is a property of the gravity_aligned_world_poses subset.

Hand annotations use a 21-joint representation and are model-estimated. Validity fields identify usable observations. Fingertip orientations are derived from terminal finger-segment orientations. Use the supplied timestamps, joint definitions, units and rotation conventions when loading pose sequences or projecting annotations onto video.

Integrity and use

Dataset curation includes checks of file integrity, stereo frame correspondence, timing, native video dimensions and annotation indexing. The release includes file lists and SHA-256 checksums for verifying downloaded content.

The data supports research in 3D hand pose, egocentric manipulation, hand-object interaction and spatial motion analysis. Select the subset whose coordinate representation matches the intended analysis.

License and contact

Handwork-100 is licensed under CC BY-NC 4.0. Noncommercial use is permitted under the license terms, including attribution to OnePiece Labs, a license link and an indication of changes.

For commercial licensing, evaluation access and custom capture, contact data@onepiecelabs.xyz.

Citation

@misc{onepiecelabs_handwork100,
  author = {{OnePiece Labs}},
  title = {{Handwork-100}},
  url = {https://huggingface.co/datasets/onepiecelabs/Handwork-100}
}

Website · Data inquiries

Downloads last month: 65

Total file size:

790 GB