Kitchens
Indian, European and Brazilian home kitchens, lived-in and cluttered, plus small commercial kitchens and pantries.
Cooking preparationWashing upCupboards and drawersAppliancesPouring and scooping
Egocentric video, depth, teleoperation and handheld-gripper demonstrations — recruited, consented and QA-reported on the same bench that has shipped four thousand hours of speech. The people on camera are doing work they actually do.
Physical-AI capture runs on the bench that has shipped 4,000+ hours of speech: the same recruiting to a locale, consent on every file, a calibration gate before paid work and a QA report on every batch. Crews record in real homes, shops, workshops, farms and light-industrial sites across nine locales, on rigs we supply or yours, and the bench grows with every program. A written spec and quote come back within 48 hours, a sample episode within five working days, and a pilot of ten to thirty episodes ships in two weeks.
The bench grows with every program: recruiting is open in all nine locales, operator crews are added per environment class as programs come in, and the fixed-term core expands with each quarter's volume.
Human video for breadth, depth for geometry, teleoperation when the embodiment has to match, grippers for grasps in rooms a robot has never seen, measured hand pose and touch for dexterity, your own rig when the stack is yours, and annotation to turn any of it into training data. Hover a card to see the capture.
First-person RGB video of real people doing real tasks.
You receive: raw video per episode, original device capture keptRGB-D, stereo and LiDAR capture for 3D scene understanding.
You receive: rgb video and aligned depth per episodeHuman-guided robot demonstrations, one episode at a time.
You receive: synchronised camera streams per episodeIn-the-wild wrist-gripper demos for grasp generalisation.
You receive: fisheye video per episodeAbsolute 3D hand pose ground truth, occlusions included.
You receive: head and wrist video per episodeContact and pressure streams for force-aware manipulation.
You receive: video plus synchronised pressure stream per episodeYour hardware and protocol, our crews, rooms and QA.
You receive: capture in your format and layoutSegmentation, captions, masks and success labels on Foundry.
You receive: labels per episode in your schemaTasks are recorded where they actually happen — not in a studio dressed to look like a kitchen. Hover a room to see the kind of setting the protocol calls for.
Indian, European and Brazilian home kitchens, lived-in and cluttered, plus small commercial kitchens and pantries.
Real Western homes in Spain, Germany, France and Italy, and apartments in India, Brazil, Japan and Korea. Sofas, wardrobes, shelves, laundry.
Open-plan and small offices: desks, pantries, meeting rooms, filing and cable clutter.
Kirana stores, groceries, pharmacies and supermarket aisles, with real stock and real customers out of focus.
Racking aisles, packing benches, courier depots and small production lines with pallets, trolleys and parts bins.
Mechanical and electronics workshops, repair benches and robotics labs where a follower rig can be set up.
Polytunnels, packing sheds, market stalls and yards, in daylight that changes from one episode to the next.
Your site, or a space recreated to your drawings, run by an NDA'd crew with limited-visibility workflows and no clip leaving your region.
A written brief, a tested crew, a pilot you review, and a weekly cadence after that — with the same report attached to every batch.
The task family, the environments, the rig, the sensors, the episode count. One page, agreed before anything is bought or anyone is recruited.
A written capture protocol — framing, lighting, where a task starts and ends, what counts as a failed take — and the kit list. Operators are trained and tested on it before a single episode is paid for.
A small batch, ten to thirty episodes, delivered with QC status and a quality report. This is where the protocol breaks — on purpose, and cheaply.
You review the pilot, we fix the protocol, then the run scales on a weekly delivery cadence with the same report attached to every batch.
Ten checks with thresholds written into the protocol. An episode that fails one is re-shot or rejected, and either way it is not billed.
The task stays in frame; hands and workspace inside the protocol's bounds for the whole episode.
No blown highlights or focus hunting beyond the tolerance set for the room.
No camera whip, no dropped frames; IMU continuous where the rig has one.
Every stream aligned within one frame; timestamps monotonic across cameras, state and action logs.
Start and end states match the SOP; no skipped or reordered steps unless the protocol allows it.
Inside the episode window for the task family; overruns are trimmed to the protocol, not the clip.
Environment, task, rig, operator id, session and consent reference present and valid on every episode.
No bystander faces, screens or documents without consent; blurred where the protocol flags it.
Intrinsics and extrinsics on file per rig and session for depth, stereo and multi-camera capture.
Where annotation is ordered, agreement against the gold set above the threshold before the batch ships.
One page is enough. A written spec and quote come back within 48 hours, a sample within five working days, and the pilot ships in two weeks.