Egocentric video for embodied AI
First-person video of people doing ordinary physical work — dishes, repairs, bedding, yard — on head-mounted cameras, in real homes and worksites. Every clip is commissioned: we meet with you, build the shot list around what your model is missing, and record to that spec.
How we work
Before any shot list exists, we want to see the failure cases — the tasks where the policy stalls, the objects it can't place, the grip it never recovers from. That conversation is what the collection gets built around.
Commission an hour. Look at it properly, on your own loader, against your own eval. If the framing, the pacing, or the level of mess isn't right, we adjust and reshoot before you commit to volume.
Labels, task naming, segment boundaries, and manifest structure follow whatever your pipeline already expects. We'd rather match your format than hand you a conversion job.
What we collect
No studio, no scripts, no set dressing. A wearer puts on the camera and does the thing they were going to do anyway, at their own pace, with their own tools. That means both hands stay in frame, objects are where people really leave them, and the clip keeps the fumbles — the dropped sponge, the screw that strips, the second attempt. Recovery behavior is usually the part that's missing from staged datasets.
01Kitchen & dishes
Hand-washing, loading and unloading, wiping down, food prep, putting away.
02Cleaning & surfaces
Glass, counters, floors, bathrooms — spray, wipe, scrub, rinse cycles.
03Laundry & bedding
Making beds, folding, sorting, hanging, changing covers and cases.
04Repair & assembly
Power and hand tools, fastening, measuring, flat-pack builds, small fixes.
05Yard & exterior
Raking, clearing, trimming, hauling, garage and shed organization.
Sample clips
Untouched apart from a resolution pass and stripped audio for the web. Hover or tap any plate to play. Full-resolution files and the JSON manifests that ship with them are in the sample set.
repair / assembly / power driver
kitchen / dishes / hand-wash
cleaning / surfaces / glass
laundry / bedding / duvet
exterior / yard / clearing
How a collection runs
A one-hour pilot batch can be shot and delivered inside a week, which is usually the fastest way to find out whether the footage is what you pictured. Larger collections run as rolling weekly deliveries, so you can start training before the set is finished.
A call or a visit, and a walk through the cases your model handles badly. That becomes a written shot list with a per-clip spec: what has to be visible, from what angle, how long, how many repeats, how much variation between wearers. You approve it before anyone puts a camera on.
Wearers are recruited for the environments and demographics your spec calls for. Everyone signs a release that explicitly covers machine-learning training and redistribution of the footage, and everyone is paid.
Head-mounted capture in the wearer's own kitchen, garage, or yard. We don't tidy up first and we don't re-shoot for neatness. The environment is the point.
Every clip is watched end to end. Bystander faces, screens, and documents are blurred on request. Labels go to your schema — task boundaries, subtask segments, hand–object contact, tool in hand, or spoken narration.
Clips plus a JSON manifest, pushed to your bucket. You flag what's short and the next batch corrects for it.
Delivery specs
Defaults are listed below. Most of it is negotiable — tell us what your loader expects and we'll match it rather than making you write a conversion step.
Start a collection
Send a task list, or just a description of the behavior your model keeps getting wrong — the second one is fine, and it's usually where the useful conversation starts. We'll set up a call, come back with a shot list, a per-hour price, and a date. The sample set goes out the same day you ask.