Get the exact training data your robot needs, within 4 weeks.
The on-demand workforce for real-world human data. Name the task, the person, and the place, and we deliver rights-cleared video plus synced motion, built to your spec.
Real-world human data, captured to your spec.
Trusted by teams building embodied AI
Explore by room & task
The same task, captured across the whole world.
Maximum place and person diversity is what makes models generalise. Switch a room or task to see the breadth.
Counters, stovetop, appliances and cupboards — the densest manipulation space in the home.
🇯🇵 Tokyo · captured to spec
View kitchen datasetsThe problem
The data physical AI needs was never online.
Robots and embodied agents cannot learn from web text. They need recordings of real people doing real tasks, with the physical motion attached. That data was never online. It has to be captured on purpose. The bottleneck is supply.
The hard part is getting the right person, in the right place, to capture the right thing, cleared for commercial use, reliably and at scale.
Our approach
A managed, on-demand workforce.
Rather than waiting for the right contributor to upload the right thing, we source and dispatch the exact people a dataset needs.
Marketplace / upload
“A listing waits for the right contributor to find it.”
- Reactive: best suited to data that already exists in the crowd.
- Less suited to specific, rare, professional, or location-bound needs.
- Coverage depends on who is already in the pool.
Motionstack
“Name the task and the person and the place. We field them and deliver the dataset.”
- Proactive: we source and dispatch the exact people, at volume.
- Targeted by task, profession, demographic, and geography.
- Sourced and dispatched to spec, drawing on a global network.
How it works
Three steps from spec to dataset.
Tell us the place and task
A task, a person type, an environment, a location, a modality, a volume, a quality bar. We turn it into a machine-readable rubric.
Our people film it to spec
We field matching contributors and they capture on a standardised rig, so every clip is consistent and comparable.
You get it in about four weeks
A loadable LeRobot dataset: media, per-frame motion, labels, consent, provenance. Consented, owned, and cleared.
More than footage
What's in every clip.
Video by default, with per-frame motion plus optional depth, 3D hand pose, and object masks. Raw or labelled, your choice. Always with the paperwork.
Real people, filmed to a standardised rig.
Camera ego-motion every frame, plus optional 3D hand pose.
Action spans, language instructions, one taxonomy.
Signed release. Owned and audit-ready.
Schema, episodes, tasks, diversity stats.
The format
Ships as native LeRobot. Auditable end to end.
Every delivery is a loadable LeRobot dataset, the format robotics labs already ingest: per-frame motion and labels in parquet, egocentric video, and language instructions, with optional depth, hand pose, and object masks. Consent, a datasheet, and a full model-and-cost record travel with it. Pull it to your storage, or via API.
See how delivery worksCoverage
50+ countries, more every week.
Diversity of place and person is the product. Here is some of where we field today.
Licensing
We own the data and license it to you.
Pricing is per QA'd hour, with an exclusivity multiplier. Choose the tier that fits the run.
Non-exclusive catalog
The default, and the lowest price. Re-licensable, so it ships now if it already exists.
Time-boxed exclusive
Exclusive use for 6 to 12 months, then it joins the catalog.
Full exclusive buyout
We forgo all future re-licensing. The top tier.
Get the real-world data your robot needs.
Tell us the task, the person, and the place. We field it from a network of 800k contributors and deliver it to spec, cleared for commercial training, in about four weeks.