Robots learn fastest from people.
We capture first-person human manipulation data at scale and deliver it the way physical AI teams need it: calibrated, annotated, ready to train.
First-person video of real people doing real work, captured on calibrated stereo and multi-camera head-mounted rigs.
Per-frame hand and head pose in one gravity-aligned world frame, so models learn how skilled hands actually move.
Metric depth from stereo on every frame. Your models get geometry, not just pixels.
Handheld-gripper demonstrations in the Universal Manipulation Interface format, ready to drop into policy training.
Every episode is one complete task, trimmed and labeled. Tell us the tasks, scenes and formats you need and we collect to that spec.
Delivered as HDF5 sensor arrays, MP4 per camera and JSON metadata with a manifest and checksums. LeRobot and RLDS exports on request. Every delivery passes our reprojection, sync and validity checks before it ships.
We also deliver multimodal, code and expert data on request.
One clip, three layers: raw first-person video, 3D hand keypoints and the reconstructed hand mesh. A public sample is on Hugging Face. The full evaluation set is one email away.
Define the spec → Capture → Process: hand pose, head pose, calibration → QC: reprojection, sync, validity → Deliver
We capture in Jakarta today and are opening Ho Chi Minh City and Bangkok next, with a team across Silicon Valley and Jakarta. Every collector is a paid, consenting participant identified only by a pseudonymous operator ID, and every site by a pseudonymous site ID, so you get diversity you can measure without personal data you don't want. Every episode is checked against the spec before it leaves our pipeline: hand and head poses reprojected onto the video, timestamps and frame counts verified, invalid frames flagged rather than filled.