First-person video you can prove is real.
Guardian captures egocentric training data across Asia and the US. Every clip is signed on the device that recorded it, checked for physical consistency, and tied to a verified, consented contributor.
Illustrative example of the record that ships with each clip.
Raw first-person hours are easy to find. Proof is scarce.
Open releases have made unlabeled egocentric footage cheap. What training teams still can't buy easily is data they can trust, audit, and learn actions from.
Footage you can't trace
When vendors pay per hour, synthetic, re-filmed, and recycled clips find their way into the pool. Bad data wastes compute and can quietly degrade a policy.
Consent you can't audit
Legal teams now ask for contributor-level consent, site permissions, and bystander handling before data reaches a training run.
Pixels without actions
Monocular video alone gives a robot little to learn from. Hand pose, depth, and camera trajectory are what carry the action signal.
What first-person capture looks like.
Real head-mounted footage from factory floors, played with the frame-level action and hand-contact labels that ship with it.
Load a stamping press
Open-dataset sampleBuild AI Egocentric-10K · factory_001 · trimmed to 12 s, re-encoded
Fold a garment
Open-dataset sampleBuild AI Egocentric-10K · factory_002 · trimmed to 12 s, re-encoded
Finish a cast part
Open-dataset sampleBuild AI Egocentric-10K · factory_003 · trimmed to 12 s, re-encoded
Fix a motor nameplate
Open-dataset sampleBuild AI Egocentric-10K · factory_005 · trimmed to 12 s, re-encoded
Real people doing real work, in the environments robots will enter.
Asia
[Countries / cities] · [#] contributors
United States
[States / cities] · [#] contributors
Finished, QA-passed hours at the depth your model needs.
3D hand pose
Per-joint keypoints from stereo or depth rigs.
Camera trajectory
6-DoF head pose from visual-inertial tracking.
Hand-object interaction
Contact start and end, grasp type, object touched.
Task segmentation and narration
Step boundaries with plain-language descriptions.
Success, failure, recovery
Outcome labels, including deliberate recovery episodes.
Choose how much proof each dataset carries.
Each tier adds evidence on top of the one before it. You see the tier of every clip in its manifest.
Declared
Contributor attestation and metadata. The industry baseline.
Verified contributor
A real, unique, consented person is tied to every session. Duplicate accounts are screened out.
Verified capture
Recorded on an enrolled device and signed at capture. Every later edit is logged and signed.
Authenticated physical capture
Motion sensors and video agree physically, and the clip passes synthetic and re-filming screens.
From first call to production data.
Scope
We agree tasks, environments, sensor signal, annotation layers, rights, and provenance tier.
Paid pilot
Minimum 20 finished hours per week with written success criteria for moving to production.
Production
Capture across our Asia and US network, with QA inside 24 hours of each session.
Delivery
Weekly batches in your format, each with a QA report and per-clip manifests.
Built by operators who have verified identity at scale.
Guardian started in identity and fraud security. We now apply that verification work to physical-world data, so training teams can trust where every clip came from.
Founded by Joseph Tan, a Y Combinator alum (S22). Offices in San Francisco, Korea, Taiwan, and the Philippines.
[Team bios: capture ops lead, ML / data lead]
Tell us what your robot needs to learn.
Thirty minutes to scope tasks, environments, signal, and volume. You leave with a pilot plan and a price per finished hour.