The view a humanoid will have
The head camera sits at eye height. Hands enter from below, parts sit at arm’s length, and the eyes land on the next part before the hand moves. A robot’s head camera sees the same scene.
The work, as the worker sees it.
neocognix records skilled workers on live production lines and labels every reach, grasp and step, so robots learn the job from people who do it for a living.
Free sample · in your format · no NDA
Three cameras, one clock
Head, wrist and cell cameras share a hardware clock and are calibrated on site. Every frame, IMU sample and label sits on one timeline, so the cell camera resolves what the head camera cannot see.
Labeled to the step
21 keypoints per hand, a 6-DoF box on every part, gaze fixations, contacts and grasp state. Each action is tied to the line of the work instruction it carries out.
In your robot’s frame
Hand motion also ships as wrist pose and grip aperture in your robot’s base frame, beside the raw streams, so a policy can train on it without another labeling pass.
Volume is no longer scarce. Skill is.
First-person datasets now reach a million hours. Few of them hold expert work, recorded on calibrated cameras and labeled to the level a policy can use.
0.30 → 0.71
Average task completion as pretraining grew from 1,000 to 20,000 hours of action-labeled first-person video, on a log-linear curve.
NVIDIA, Feb 2026(opens in a new tab)≈2×
Performance on generalization tasks seen only in human video, once the model had diverse robot pretraining behind it.
Physical Intelligence, Dec 2025(opens in a new tab)201 vs 43
Demonstrations recorded per hour with a wearable rig, against a teleoperation setup on the same tasks.
Tao et al., May 2025(opens in a new tab)
What one episode contains
Nine streams on one clock, from raw pixels to task steps. Choose the layers you train on and we label only those.
VideoHead, wrist and cell
IMUAccel and gyro, head and wrist
Head pose6-DoF trajectory, gaze if fitted
Hand pose21 keypoints per hand
Objects6-DoF pose, boxes, track IDs
ActionsVerb and object, per hand
ContactsTouch and release events
StepsAligned to your SOP
NarrationPresent-tense sentences
A pilot with fixed edges
One site and one task family, delivered in the formats you train on. You run it through your own evals before anything scales.
- Sites
- OneA working cell, recorded during normal shifts.
- Task family
- One, chosen with youKitting, assembly, cable work, inspection…
- Recorded hours
- 50Across 6 to 10 skilled operators.
- Formats
- Any two of fiveLeRobot v3, RLDS, MCAP, HDF5, WebDataset.
- Reports
- QA and provenanceMeasured values against every published threshold.
- Time to first delivery
- 8 weeksFrom kickoff to labeled data in your bucket.
Consent you can audit
Every clip carries its own record: who recorded it, under what consent, what was blurred, and how the operator was paid.
Operators are briefed in their own language, paid for every recorded hour on top of their wages, and can pull a clip for 30 days. Faces are blurred in every view before anything leaves the site.
{
"clip": "ep_0412",
"site": {
"id": "site_027",
"agreement": "SA-027, signed 2026-07-14"
},
"operator": {
"id": "op_8f31",
"trade": "assembler, 11 years",
"briefing": "v3, operator's first language",
"purpose_disclosed": "training robot manipulation models",
"consented_at": "2026-07-15T02:32:11Z",
"withdrawal_open_until": "2026-08-14",
"paid": "per recorded hour"
},
"deidentified": {
"faces_blurred": "all views, operator included",
"screens_masked_at_capture": 1,
"screens_blurred_before_upload": 2,
"audio_during_take": "none"
},
"license": "commercial, per your MSA",
"exclusive_until": "2027-07-31"
}From the blog
Common questions
- What does neocognix do?
- neocognix records skilled workers doing real jobs with head, wrist and fixed cameras on one clock, and labels every reach, grasp and step. Robot and humanoid teams use the data to train manipulation policies.
- What is in one episode?
- Synchronized head, wrist and cell video, IMU, 3D hand pose with 21 keypoints per hand, object poses, actions, contacts, the task step each action belongs to, and narration. You choose the layers you train on.
- Which formats do you deliver?
- LeRobot v3, RLDS, MCAP, HDF5 and WebDataset, with calibration, QA and provenance files in the same folder.
- How is quality measured?
- Every delivery reports camera sync, calibration error, 3D keypoint error and label agreement against thresholds we publish on the datasheet. Episodes that fail are dropped, not patched.
- How do you handle consent?
- Operators are briefed in their own language, paid for every recorded hour on top of their wages, and can withdraw a clip for 30 days. Faces are blurred in every view, and each clip ships with a consent record.
- How do we start?
- Request a free sample episode in your format. If it holds up in your evals, a pilot covers one site and one task family, with a first delivery in 8 weeks.
Start with one episode.
We send a sample episode with every stream and label layer, in the format you train on. No NDA needed.