✦ Real-world data for Physical AI

Data that teaches robots real work.

First-person human work, labeled in full 3D context, delivered training-ready.

For humanoid, mobile-manipulation & VLA teams

30K+
operators
100+
tasks
80+
environments
10+
domains
The gap

Most robot datasets log hours, not understanding.

A VLA needs context: where, what, why, and what happens when it fails.

What's in every episode

Four questions, answered on every frame.

01WHERE

Spatial grounding

6-DoF head pose + 3D hand pose (21 keypoints), grounded with IMU.

02WHAT

Interaction

Object tracking + contact events.

03WHY

Intent

Step, goal, and full-sequence labels.

04WHAT-IF

Recovery

Slips, corrections, and retries kept in, not deleted.

Capture, locked across the fleet
1080p30fpsdFOV >120°IMU ≥200Hzsync <33msQC-gated
What sets us apart

Three things built into every dataset.

01 / RECOVERY

Failures stay in

When a grip slips, a part drops, or a step gets redone, we keep it in the footage and label it, so your policy learns to recover, not just to succeed.

02 / PROVENANCE

Every clip is accountable

Consent-first sourcing, fair pay, PII redaction, and full chain-of-custody.

03 / SCALE

Real people, real range

30K+ operators across 100+ tasks, 80+ environments, and 10+ domains, not the same lab loop repeated.

Ships in your format
LeRobotRLDSHDF5Zarr
Consent-firstPII redactionFull provenanceClear licensing

Tell us where your policy breaks.

We'll go capture it.