Physical AI · Robotics & humanoids
Teach robots the tasks people do every day.
First-person video of real work, broken into the actions a robot has to learn, with hands, objects and steps marked up.
Lab demonstrations don't look like real work.
A few tidy tasks recorded in a lab leave robots unprepared for cluttered, busy floors that change every day. Manipulation models need the real thing, at volume.
Data samples
Check our data samples.
Action labels
Hand and object marks
Fine manipulation
Bounding boxesSupported data types
What we capture and label.
- First-person video
- Wrist-camera video
- Action labels
- Hand and object marks
- Bounding boxes
- Segmentation masks
Proven in production
The standard every hour is held to.
- 1.8M
- Frames annotated in six months for one robotics lab
- 100%
- Of recorded hours reviewed by a person
- 98%
- Labelling accuracy with human-in-the-loop checks
- 4
- Consent languages: Arabic, English, Filipino and Urdu
Other physical AI verticals
What kind of data do you capture?
First-person video of real work: picking, packing and shelf stocking in warehouses, kitchen work, housekeeping, laundry and cleaning, and hospital portering. We also capture photographs of handwritten Arabic pages.
How is it recorded?
Hands-free, with a phone sideways on a head strap, the Roboband headband (up to eight synced cameras) or the wristband camera. Phone recording works on iPhone 11 to 17, Pixel 6 to 9, Galaxy S21 to S24 and more.
How do you handle consent and privacy?
Every worker agrees before recording, in Arabic, English, Filipino or Urdu. We never process content to identify anyone.
How do you check quality?
A person reviews every recorded hour. Only accepted hours are paid for and delivered, and labelling is checked with a human in the loop.
Bring real-world data to your robotics & humanoids work.
Tell us what your robot needs to learn. We will scope the environments, tasks and labels with you.
