Human data for physical AI

The internet taught AI what humanity says.We teach it what humanity does.

First-person video of people at work, processed into 3D hand pose, depth and labels for training robots.

Request data Join the network

[01] Why now

Robot models are ready. The data isn't.

of human video went into pretraining NVIDIA's GR00T N1.7: 20K hrs
robot task success from pretraining on human video (EgoScale): +54%
invested in robotics and physical AI in 2025: $27.6B

Sources: NVIDIA, Jun 2026 · NVIDIA EgoScale, Feb 2026 · PitchBook, 2025

[02] Egocentric data

Filmed from behind the eyes.

The side view shows the wearer. The head view is what their camera records. That's how the footage above gets filmed.

About the footage

Devices: Stereo and mono head-cams, UMI grippers, a lab's own rigs
Sensors: Full IMU, plus LiDAR or depth where the rig allows
Settings: Homes, offices, industrial sites, public spaces
Processing: 3D hand pose, camera trajectory, metric depth, VLM labels
Privacy: Faces removed and PII scrubbed before delivery
Verification: Verified, provenance tracked
Consent: Explicit, given by contributors
Rights: We ask capture teams to confirm footage is rights-cleared

The camera moves with the head.

Tip your head down to pour and the camera tips with it, toward the mug and bowl.

Things get dropped.

Someone fumbles the bowl and catches it. That's in the footage too, and a robot needs to see it.

Who gives consent, and who clears rights?

Contributors give explicit consent. We ask capture teams to confirm their footage is rights-cleared and unencumbered.

Which sensors come with the video?

It depends on the rig. Sensors can include IMU, LiDAR or depth. Ask us what a specific rig records.

How does this become robot training data?

We fine-tune vision-language-action models on it, alongside demos operators record by driving real robots.

[03] Built to spec

Every lab asks for different data.

Labs set the task, the setting and the rig. Our capture teams go and record it.

Egocentric video

Stereo and mono head-cams, with full IMU.

UMI grippers

Handheld grippers that record the grasp a robot will make.

Custom rigs

A lab's own cameras and sensors, run by our teams.

Teleoperation

Operators drive real robots through the same tasks.

[04] Capture

Who holds the camera.

Our capture teams do the filming. We work with them as partners.

[05] Output

Every clip comes back in 3D.

Our in-house pipelines turn raw footage into data a robot can act on.

[01]
3D hand pose
Every finger joint, every frame.
[02]
Camera trajectory
Where the head moved, in metres.
[03]
Metric depth
Measured with stereo cameras.
[04]
Labels
Captions, actions and objects from VLMs, then verified.
[05]
Privacy
Faces removed, PII scrubbed, before anything ships.

[06] Pipeline

The camera is step one.

Capture
Egocentric video, with IMU, LiDAR or depth where the rig allows.
Annotate
VLMs label the footage. We add 3D hand pose, depth and the camera path, and remove faces. All on NVIDIA GPUs.
Teleoperate
Operators drive real robots to record demonstrations.
Train
We fine-tune vision-language-action models on the footage and demos.

[07] Built on NVIDIA

All our processing runs on NVIDIA GPUs.

NVIDIA GPUs: Every in-house pipeline: 3D hand pose, trajectory, depth, PII scrubbing and labels.
Isaac GR00T N1.7: Fine-tuned on our human and robot demos. Datasets ship in its LeRobot format.
Cosmos Reason 2, Curator: Captioning and curation inside our processing.
Physical AI Data Factory: Real human footage as the seed for curation and augmentation.
Isaac Teleop: Teleoperation capture on real robots.

[08] Teach

The arm copies the hand.

We turn human demonstrations into robot-ready training data.

[09] For robot labs

We sell processed data directly to robot labs.

We don't build robots, so we never compete with the labs we supply.

Datasets

Processed, verified footage, licensed per hour.

Custom collections

Your tasks, devices and settings, recorded to spec.

Robot demonstrations

Teleoperation data, joined with the human video.

[10] Work with us

Buying data, or filming it?

For robotics & embodied-AI labs

Request data.

Email or call with the task and volume. We don't compete with our customers.

For capture teams

Join the network.

Pick an AI. It asks 9 questions and drafts your email and WhatsApp message to us.

ChatGPT Claude Gemini Mistral

What the 9 questions cover

Rights
Devices
Sensors
Archived volume
New volume / month
Pilot speed
Pricing
Domains
Geography