OnePiece Labs

Handwork-100-Preview on @huggingface: 25 episodes, 30 minutes, approval is automatic.

Activities include object organization, flower arrangement, device repair, screw driving, planting, painting, folding clothes, block building and laboratory tasks, each recorded on a calibrated stereo rig with per-frame hand and head pose.

Original video + HDF5, or prepared LeRobot v3 exports: huggingface.co/datasets/onepi…

#EgocentricData #PhysicalAI #RobotLearning #OpenData

Sorting metal parts by hand, seen from the collector's own view. Both hands work inside the tray at once, and one often covers the other, which is the kind of contact a robot has to learn from people.

The 21-joint hand skeleton is an overlay shown for illustration.

Open preview: huggingface.co/datasets/onepi…

#PhysicalAI #EgocentricData #HandPose #RobotLearning

Our capture operation is our own. Trained collectors on calibrated rigs, working in real homes, shops, kitchens and workshops 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. Every delivery ships with the calibration and checks needed to trust it.

onepiecelabs.xyz

#PhysicalAI #EgocentricData #Robotics #DataCollection

Handwork-100 ships in three coordinate representations so you can train in the frame your model expects:

• camera_frame_poses: hand poses in camera coordinates
• world_frame_poses: hand and head-mounted camera poses in world coordinates
• gravity_aligned_world_poses: the same, in a gravity-aligned world frame with Z up

The 30-minute preview on @huggingface carries about 10 minutes of each: huggingface.co/datasets/onepi…

#PhysicalAI #EgocentricData #RobotLearning #Robotics

What is inside one Handwork-100 episode:

• Synchronized stereo video, 30 fps, per-frame exposure timestamps
• 21-joint skeleton for each hand, position and orientation per joint, plus wrist pose
• 6-DoF head pose in a gravity-aligned world frame
• Intrinsics, distortion and stereo extrinsics for every camera
• Validity flags on every row: occluded joints are marked, not guessed
• Task label and action segments, one complete task per episode

Delivered as HDF5 + MP4 + JSON with manifest and checksums; @LeRobotHF and RLDS exports on request. Dataset card: huggingface.co/datasets/onepi…

#EgocentricData #PhysicalAI #RobotLearning #HandPose