For robotics and embodied AI teams
Robots learn from hands at work. We record them.
Contributors across Brazil strap a phone to their head and do real tasks at a natural pace. You get first-person footage of the physical world, with the variety lab data doesn’t have.
Request sample clipsShared under NDA once we have your spec
TASK: CHOPPING VEGETABLESSETTING: HOME KITCHENDataset spec
What is in a delivery.
Defaults below. Tasks, settings, clip length and volume are set per project and written into the order.
- Viewpoint
- First person, phone mounted on the head. Stereo head rigs on request
- Framing
- Wide angle, both hands in frame, natural pace
- Resolution
- 1080p minimum (1920 × 1080), ultra-wide 0.5x lens
- Frame rate
- 30 fps constant (60 fps on request)
- Clip length
- 2 to 30 minutes per recording, set per task. Delivered as approved segments, cut to your spec: whole task, per action or fixed length
- Settings
- Kitchens, workshops, construction sites, farms and homes
- Metadata
- Task label, setting, hands visible, light check, review status
- Files
- MP4, H.264 plus JSON metadata per clip
- People in frame
- Only the contributor appears
- Consent
- Signed per contributor
- Monthly capacity
- Set after a paid pilot. Pilot batches of 100 to 200 h
- Lead time
- 7 days from brief to first batch
Tasks and settings
Real work, in the places it happens.
Recorded by the people who do the job every day, from nurses to mechanics, cooks and farmers.
HomeDishes, laundry, cooking, cleaning10 tasks
Food workPrep, baking, packing meal boxes5 tasks
Repairs and upkeepPlumbing, sockets, painting, car and moto care7 tasks
Sewing and craftsSewing, cutting fabric, embroidery, crafts6 tasks
Shops, farms and cleaningRestocking, order packing, gardens, harvest, site cleaning9 tasks
37 tasks in the catalog today, and any other task on demand. Send the scenario your model is missing: we write the brief, recruit the right people and pay them to record it.
Review and privacy
Every clip is checked before it leaves.
Homes and worksites contain things you don’t want in a training set: bystanders, documents, screens, addresses. The pipeline is built to catch them.
- Step 1CaptureContributor follows the task brief and framing guide.
- Step 2UploadThrough Verakin, tied to a verified identity and consent.
- Step 3Automated checksLight, framing, hands in view, motion blur. Faces and screens flagged.
- Step 4Human reviewLight, framing and task completion reviewed per clip.
- Step 5RedactionFaces of other people, screens and documents are blurred. Clips with children are rejected.
- Step 6DeliveryApproved clips with labels and metadata.
100%of clips reviewed by a person before delivery
100%of clips linked to a signed consent record
48 hto replace a clip you reject, at no cost
1 fileof JSON metadata per clip, schema-checked
- Model training
- Perpetual and worldwide, for training and evaluating models
- Commercial use
- Allowed in commercial models and products
- Exclusivity
- Exclusive for custom collections, non-exclusive for catalog data
- Contributor consent
- Signed per contributor · AI training, evaluation and sublicensing
- Retention and deletion
- Kept 12 months for re-delivery, then deleted unless the contract says otherwise
See the footage before you commit.
Send us your spec: the tasks and settings you need. After a mutual NDA, we send sample clips with their metadata and a project quote.