OnePiece Labs

Assembling small plastic parts from two bins. Each cycle is a pick, an alignment and a press fit, repeated with both hands, and a first-person camera sees every step of it. The 21-joint hand skeleton is an overlay shown for illustration. Open preview: huggingface.co/datasets/onep… #PhysicalAI #EgocentricData #Manipulation #RobotLearning

Beyond head-mounted capture, we collect handheld-gripper demonstrations in the Universal Manipulation Interface (UMI) format, built to the customer's task list and scenes. On request we also supply multimodal, code and expert data for teams that need more than video. Scope a collection: [email protected] · onepiecelabs.xyz #PhysicalAI #UMI #RobotLearning #TrainingData

Squaring a deck of cards takes two hands, fine finger control and constant small corrections. Thin, flat objects like these are among the hardest things for a robot hand to pick up and align. The 21-joint hand skeleton is an overlay shown for illustration. Open preview: huggingface.co/datasets/onep… #PhysicalAI #EgocentricData #Dexterity #RobotLearning

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/onep… #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/onep… #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/onep… #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/onep… #EgocentricData #PhysicalAI #RobotLearning #HandPose

OnePiece Labs is a physical AI data company. In real workplaces, we record first-person video of real people doing real work on calibrated stereo head-mounted rigs, and deliver metric-scale 3D hand pose, head pose, depth and task labels that robotics teams can train on directly. Start with Handwork-100-Preview, a 30-minute open sample on @huggingface: huggingface.co/datasets/onep… Full dataset and custom collection: onepiecelabs.xyz #PhysicalAI #EgocentricData #Robotics #TrainingData