Spaces:
Running
Running
| title: README | |
| emoji: 🌍 | |
| colorFrom: green | |
| colorTo: blue | |
| sdk: static | |
| pinned: false | |
| <p align="center"> | |
| <img src="https://huggingface.co/spaces/onepiecelabs/README/resolve/main/assets/banner.png" alt="OnePiece Labs" width="100%"> | |
| </p> | |
| ## About us | |
| **Precision egocentric data for physical AI.** | |
| Robots learn fastest from people. OnePiece Labs captures first-person video of real people doing real work, on calibrated stereo and multi-camera head-mounted rigs, and delivers metric-scale 3D hand pose, head pose, depth and task annotations that robotics and physical AI teams can train on directly. We also collect handheld-gripper demonstrations in the Universal Manipulation Interface format. | |
| We run our own capture operation in Southeast Asia: Jakarta today, Ho Chi Minh City and Bangkok next. Every person on camera is a paid, consenting participant, every episode is one complete task, and every delivery ships with the calibration and checks needed to trust it. Our customers are robot foundation model teams and research labs in the US and Europe. | |
| [Website](https://onepiecelabs.xyz) · [data@onepiecelabs.xyz](mailto:data@onepiecelabs.xyz) · [X](https://x.com/kjia562) · [LinkedIn](https://www.linkedin.com/company/onepiecelabsxyz) | |
| ## Capture and annotation capabilities | |
| The capabilities below describe our capture and delivery offering. Modality coverage and coordinate conventions are specified in each dataset card and project brief. | |
| - **Synchronized stereo video**, H.264 at 30 fps, with per-frame exposure timestamps and one clock per episode | |
| - **21-joint hand skeleton for each hand**, position and orientation per joint, plus wrist pose | |
| - **6-DoF head pose** from SLAM and IMU, in a fixed, gravity-aligned world frame set once per episode | |
| - **Calibration delivered**: intrinsics, distortion and stereo extrinsics for every camera | |
| - **Validity flags on every row**, so missing or occluded joints are marked instead of guessed | |
| - **Task label and action segments** with start and end times; successful outcomes only, setup and retries trimmed | |
| - **Pseudonymous site and operator IDs**, so you can balance diversity without seeing anyone's identity | |
| Delivered as HDF5 + MP4 + JSON per episode with a manifest and checksums. LeRobot and RLDS exports on request. Depth maps and point clouds are available for projects that need them. | |
| ## Handwork-100 | |
| Our first named collection: **100.64 hours of egocentric stereo recordings across 20 scene categories, captured by 30+ collectors.** | |
| <table style="display:table;width:100%;table-layout:fixed;"> | |
| <thead> | |
| <tr> | |
| <th width="50%"><a href="https://huggingface.co/datasets/onepiecelabs/Handwork-100-Preview">Explore Preview</a></th> | |
| <th width="50%"><a href="https://huggingface.co/datasets/onepiecelabs/Handwork-100">Request Full Dataset Access</a></th> | |
| </tr> | |
| </thead> | |
| <tbody> | |
| <tr> | |
| <td valign="top"><strong>30-minute sample</strong><br>Sign in and request access; approval is automatic.</td> | |
| <td valign="top"><strong>Complete 100.64-hour dataset</strong><br>Sign in and request access; manual approval is required.</td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| It comes in three subsets so you can pick the frame you train in: | |
| | Subset | Hand pose | Head pose | Frame | | |
| |---|---|---|---| | |
| | `camera_frame_poses` | per frame | — | camera | | |
| | `world_frame_poses` | per frame | per frame | world | | |
| | `gravity_aligned_world_poses` | per frame | per frame | gravity-aligned world, Z up | | |
| The complete Handwork-100 collection has been uploaded and passed file-integrity verification. Its Dataset Card and file listings are publicly visible on Hugging Face; access to dataset contents requires manual approval from OnePiece Labs. Sign in to Hugging Face and request access on the [Handwork-100 dataset page](https://huggingface.co/datasets/onepiecelabs/Handwork-100). For commercial licensing, contact [data@onepiecelabs.xyz](mailto:data@onepiecelabs.xyz). | |
| A [30-minute Preview](https://huggingface.co/datasets/onepiecelabs/Handwork-100-Preview) is now available, with approximately 10 minutes from each coordinate subset. Preview access is approved automatically after you sign in and submit a request; access to the full Handwork-100 collection requires separate manual approval. Prepared observation-only LeRobotDataset v3 Preview exports and [self-service conversion tools](https://huggingface.co/datasets/onepiecelabs/Handwork-100-Preview/blob/main/lerobot/README.md) are also available. The full native release can be converted using the pinned workflow; a preconverted 100-hour LeRobot copy is not included. | |
| ## How a project runs | |
| 1. **Define the spec.** Tasks, scenes, cameras, frame conventions and delivery format, written down before capture starts. | |
| 2. **Capture.** Trained collectors on calibrated rigs, in real homes, shops, kitchens and workshops across Southeast Asia. | |
| 3. **Process.** Hand pose, head pose and calibration computed per frame; episodes trimmed to one task each. | |
| 4. **Check.** Reprojection error against the video, stereo sync, validity coverage and diversity targets, per episode. | |
| 5. **Deliver.** Batches on a schedule, with a manifest, checksums and a changelog. | |
| ## Work with us | |
| We take custom capture briefs: your tasks, your scenes, your format. Write to [data@onepiecelabs.xyz](mailto:data@onepiecelabs.xyz) for evaluation access to Handwork-100 or to scope a collection. | |