Wear → demonstrate → check → approve
We capture the same task repeatedly across different body profiles and different settings. We publish the criteria and the reasons for rejection because data is only worth something if a buyer can verify its quality themselves.
The way in
WearWEAR
You pick the rig that suits the task, from the three tiers T1, T2 and T3. Consent is taken before the rig goes on. The scope of that consent is recorded: whether commercial training is allowed, third-party transfer, and the terms of withdrawal.
DemonstrateDEMONSTRATE
Sessions run in 30-minute blocks. The task sheet states the success conditions and the edge cases. The same task is collected repeatedly, across different body profiles and different environments.
QAQUALITY GATE
A session that fails the hand-tracking confidence check is rejected. Every frame is tagged with pose_source (0 measured · 1 monocular estimate · 2 interpolated · 3 synthetic), so the provenance of quality is never hidden.
LicenceLICENSE
You choose between compute-to-data and export. DRM on training data is not possible, so we trace leak paths through the contract and a per-customer fingerprint.
What we check before approving
All five must pass for a session to be approved. We always share the reason for a rejection with the contributor, and rejections caused by equipment faults are not deducted from payment.
| Check | Pass condition | Why |
|---|---|---|
| Wrist SE(3) continuity | Dropouts under 5% of the session | The first thing a buyer looks at. Where it breaks, that segment is unusable for training. |
| Thumb–index aperture | 0.2s either side of the grasp is all measured | If the moment of grasping is missing, the data is meaningless as manipulation data. |
| Occlusion | Both hands hidden at once for under 2 seconds | The most common failure mode of camera-based tracking. |
| Pose source grade | pose_source recorded on every frame | If measured and estimated values are mixed, a buyer cannot verify quality. |
| Consent scope match | Nothing captured beyond the consented items | If it goes beyond, we discard the whole session. No exceptions. |
We do not mix measured and estimated values
Every frame carries this value. Buyers can filter out estimated segments before training, and we can put a number on quality. Mixed together, neither is possible.
- 0Measured
arkit_measured
Measured directly by the device
- 1Estimated / generated
monocular_est
Estimated from monocular video — depth error 5–11cm
- 2Estimated / generated
interp
Interpolated from neighbouring frames
- 3Estimated / generated
synthetic
Procedurally generated — this is what the viewer samples are
Every motion currently shown in the viewer is pose_source=3 — a procedurally generated synthetic sample, not captured data.
Thirty minutes is one unit
A session is 30 minutes: 5 for wearing and calibration, 20 for the demonstration, 5 for review. Run it longer and IMU drift accumulates while concentration drops, so quality degrades towards the end.
We do not delete failed attempts. Dropping an object and picking it up again is exactly the recovery behaviour a robot needs to learn, which makes it worth more than a set of clean successes. That is why the task sheets list edge cases separately.
We repeat the same task across body profiles and settings. Fifty hours of one person washing dishes in their own kitchen is worth far less than fifty hours of seven body profiles washing dishes in seven different kitchens.
Rig specifications →Check it in the 3D viewer →Become a contributor →