Frontier DataPhysical AI
Human data for robots that do skilled work.
First-person video, skilled demonstrations and expert labelling of real, specialised tasks, collected with consent across industries.
Instruction: "Fasten the bracket with four M3 screws, diagonal order."
Three ways we collect it.
Every dataset is built from people who do the work for a living, recorded on the job or in a matching setup.
- 01
First-person capture
Head and chest cameras on skilled people doing their real work, with hands, tools and parts in view.
Egocentric video · depth · IMU
- 02
Skilled demonstrations
Experts perform the task step by step for imitation learning, including remote-controlled runs on your robot or a matching rig.
Teleoperation · gripper demos · retries kept
- 03
Expert labelling
People who do the work label it: steps, tools, grasps, mistakes and recoveries, with a written instruction for every episode.
Steps · hands and tools · success and failure
Skilled tasks across industries
We recruit people who do the work every day, so your robots learn from experts, not stand-ins.
Manufacturing
Assembly, fastening, finishing, quality inspection, packing
Logistics
Picking, sorting, kitting, packing, loading
Food and kitchens
Prep, cutting, cooking, plating, cleaning down
Home and hospitality
Folding, tidying, making beds, room turnover
Labs and healthcare
Sample handling, pipetting, instrument prep
Need a task that isn't listed? Tell us the robot and the task, and we'll find the people who do it.
Ethical by design
The people in the data come first.
Recording someone at work is personal. These rules apply to every dataset we deliver.
- Consent, in person
- Asked directly from every person recorded, in their own language. Never only through an employer.
- Fair pay, paid directly
- Every session is paid, straight to the person who did the work.
- Never used to track work
- Recordings are never used to measure productivity or rank workers.
- Blurred by default
- Faces, screens and bystanders are blurred. Audio is removed unless the task needs it.
- Free to withdraw
- Anyone can withdraw, and their recordings are left out of every future delivery.
Train-ready, not just recorded
Checked, synced and delivered in the formats your stack already uses.
- Formats
- MCAP, RLDS, LeRobot, ROS bag, MP4, JSON
- Capture
- Head-mounted RGB or stereo, wrist cameras, IMU; or your own rig
- Labels
- Task steps, written instructions, hand keypoints, objects, failures and retries
- Quality
- Every clip checked for visibility, sync and task completeness; experts review the labels
- Documentation
- A dataset card with tasks, settings, worker skill levels and consent records
- Delivery
- To your AWS, Google Cloud or Azure bucket
Tell us the task. We'll find the hands.
Share your robot, task and format. We'll scope a sample dataset with you.
Or email us at contact@kodiva.ai