Factory work,
in first person.
Skilled operators wear Ray-Ban Meta glasses and head cameras through real shifts at machining and die-casting plants. We turn what they see and do into labeled, checked training data for robots.
ExampleFaces blurred and consent on file before this clip left Prep.
Download sampleRecorded the way
robot labs record.
The rigs and methods robot-learning teams use for first-person data are our standard: a head camera with motion sensors, both hands in frame, every task split into ordered steps, delivered in LeRobot. Every lab we work with gets the same spec.
Request a sampleCapture
- Rig
- A head-mounted camera with motion sensors, worn at eye height. Ray-Ban Meta glasses where the lightest kit matters. Every episode
- Video
- 1080p or better at 30 fps or above, wide field of view, both hands in frame Every episode
- Motion
- Head IMU on head-camera rigs, recorded with the video and synced to its clock in Prep Every episode
- Calibration
- Camera intrinsics per rig from a checkerboard, shipped with the dataset On request
Structure
- Episodes
- Cut where the station changes, each with its task, station ID and step Every episode
- Steps
- Every task split into the ordered steps of the plant's own procedure Every episode
- Off task
- Walking and waiting between stations marked off station, never counted as work Every episode
- Hand pose
- 3D hand pose for both hands, reconstructed from the video On request
Release
- Consent
- Signed by every wearer in their own language under India's DPDP Act Every episode
- Privacy
- Faces, customer parts and paperwork blurred before footage leaves Prep Every episode
Delivery
- Format
- LeRobot v3.0, video and sensor streams on one timeline Every episode
- Report
- The full validation report with every dataset, warnings included Every episode
Every episode: built into how we capture and process today. On request: scoped into a pilot when a lab needs it.
What every
episode carries.
The same rig and the same pipeline on every floor, so a lab can train on footage from one plant and test on another.
Read the standard- Wearer
- A skilled worker on a normal shift
- Camera
- Head-mounted camera with motion sensors, or Ray-Ban Meta glasses, 1080p or better at 30 fps
- View
- First person, head-mounted, both hands in frame
- Pace
- Production speed, real parts, real cycle times
- Labels
- Station and step on every episode, read from a printed tag at each station
- Output
- Episodes in LeRobot v3.0
public robot-learning datasets we checked were recorded on a working production line.
See the 18 we checked
| Dataset | Where it was recorded |
|---|---|
| Ego4D | Homes, outdoors, some real workplaces (trade shops, farm and garden work). A minority of its 3,670 hours. |
| Ego-Exo4D | Real kitchens, bike shops, gyms and pitches |
| EPIC-KITCHENS-100 | 45 home kitchens |
| HD-EPIC | 9 home kitchens |
| EgoDex | Tabletop tasks with household objects |
| HOT3D | Lab mock rooms |
| ARCTIC | Motion capture lab |
| HOI4D | 610 indoor rooms |
| Assembly101 | Lab, toy vehicle assembly |
| MECCANO | Lab, toy motorbike assembly |
| IndustReal | Industrial-style lab, toy car assembly |
| ENIGMA-51 | Industrial-style lab, staged board repair |
| HA-ViD | Lab assembly workbench |
| IndEgo | University research labs and test field |
| DROID | 564 scenes in labs, offices and homes |
| BridgeData V2 | Toy kitchens and lab |
| Open X-Embodiment | 60 datasets from research labs |
| AgiBot World | A purpose-built collection site with home, retail and industrial mockups |
Checked October 2026 against each paper or project page. Ego4D and Ego-Exo4D include some real paid work (trade shops, bike shops, farm and garden work), none of it on a production line. The other 16 are homes, labs or purpose-built mockups.
How a shift
is captured.
Worn by the people who do the work, at the speed they do it.
Glasses, not a rig.
Ray-Ban Meta glasses, or a light head camera worn with the plant's own safety glasses. Nothing that slows the line or changes how the job is done.
- A head camera with motion sensors, or Ray-Ban Meta glasses
- Worn with the plant's own safety glasses
- Both hands in frame from the wearer's eyes
The job, at cycle speed.
Workers record the job they already do, at cycle speed, with the slips and recoveries that staged data never has.
- Real parts and real cycle times
- Slips and recoveries kept, not cut
- Split into episodes at every station change
Signed before anyone records.
Every wearer signs in their own language under India's data protection law before recording. Faces, customer parts and paperwork are blurred before footage leaves Prep.
- Consent in the wearer's own language
- Recorded under India's data protection law
- Faces, customer parts and paperwork blurred in Prep
Everyone on the line is paid.
Operators are paid for every capture hour that passes the checks, and the plant earns a share of every licence.
- Operators paid per hour that passes the checks
- The plant earns a share of every licence
- Nothing about the job changes
ExampleHow every step
gets its label.
From the floor, not from someone guessing at the video afterwards.
- 01
A tag at every station
Each station gets a printed marker at eye level. The tag is just a number; one file per plant says which station and which step of the plant's procedure that number means.
- 02
Read from the footage
When the wearer settles at a station, the tag in view starts a new step. A hand over the tag or a glance away does not break it, and a neighbour's tag across the floor is ignored.
- 03
Split at the change
Every clip is cut where the station changes. Each episode carries the step and the station ID, and walking between stations is marked off station rather than counted as work.
- 04
Labels that keep
Each clip keeps a copy of the station map it was recorded under, so a station that changes jobs next month never relabels last month's footage.
Where robot
data comes from.
Three sources labs already use, and the one that is missing.
Acted out for the camera
Paid crowds repeat a task at home. Clean motions, nothing at stake, no production pace.
Precise and narrow
Exact robot actions, but one rig in one lab, and every hour is expensive.
Cheap and close
Scales without limit. Contact, friction, cables and soft parts never quite match the floor.
What we record
Skilled people at cycle speed on real production lines, seen through their own eyes, with the slips and recoveries a policy has to learn.
Every upload
is checked.
Before anything ships. The report goes to the lab with the data, warnings included.
- Format
- LeRobot v3.0, v2.1 or RLDS detected and its metadata parsed
- Episodes
- Every declared episode present, frame counts match the data files, no corrupt files
- Video
- Every video opens and decodes with ffprobe, at the frame rate it claims
- Sync
- Video and sensor timestamps checked for drift against each other
- Labels
- Share of episodes carrying a task label
- Faces
- Sampled frames screened for faces. Flagged faces are blurred or covered by signed consent before publishing
Tell us what your
robots need to learn.
A sample comes before you commit to hours.
- 01Tell us the tasks
What the robot has to learn, the setting and how many hours.
- 02Get a sample
We put glasses on a line that does the work, record a sample and run it through the checks.
- 03Scale it up
Order the full set, or an ongoing feed of new episodes from more floors.
Run a factory?
Get paid for the work your line already does.
We bring the cameras, handle consent and blurring, and sell to the labs. Operators are paid per recorded hour and the plant earns a share of every licence. Nothing about the job changes.
How the partnership works