First-person video that survives contact with a policy.

Egocentric video is footage recorded from a head-mounted camera worn by the person performing a task, so the camera shares the actor’s viewpoint and moves with their head. For embodied AI it matters because the observation distribution matches what a robot with a head or chest camera actually sees: hands entering frame from below, occlusion during grasp, motion blur during reach, and a first-person view of the manipulated object. Third-person footage does not contain that geometry, so policies trained on it transfer poorly.

NVIDIA’s EgoScale work reports a log-linear relationship between egocentric hours and downstream policy success out to 20,854 hours with no observed saturation, so more hours keep helping. In practice the binding constraint is fitness, not volume: teams typically discard around 90% of bulk footage as unusable for manipulation. A useful starting point is 50–100 validated hours against one named skill, which is enough to measure whether your architecture responds to the data before you commit to a thousand hours.

Ego4D’s license does permit commercial use, so the honest gap is fitness rather than legality: it is roughly 3,670 hours of unstructured daily activity, much of it with heavy motion blur, no manipulation-grade 3D hand pose, and no frame-accurate action labels. EgoDex has excellent hand pose but is licensed CC BY-NC-ND, which genuinely rules out commercial training use. Firsthand captures against a named skill spec, hardware-synchronizes every stream to under two milliseconds, ships depth, 3D hand pose, body pose, segmentation and action segments on every clip, and attaches a commercial buyer-owned license with a documented consent chain.

A head-mounted global-shutter 4K/60 camera with a 120° lens, two wrist-mounted 1080p/60 cameras, an active-stereo depth sensor at 848×480/30, and 200 Hz 6-DoF IMUs. All streams are hardware-triggered onto a single PTP clock domain, and intrinsics and extrinsics are re-solved with a checkerboard target at the start and end of every session so drift is detectable rather than assumed.

We quote per engagement rather than publishing a rate card, because the price is driven by things we cannot guess from a web page: which environments you need, how thin the coverage is in them, how dense the annotation has to be, and whether you want exclusivity. The sample pack is free and ungated, so you can evaluate the data before any commercial conversation. Tell us what you are training and your target hours in the intake form and you will get a scoped estimate — hours, price, delivery date, and the coverage cells we would need to fill — within one business day. Billing is always per validated hour: footage that passed sync and QA against your spec. Rejected episodes, recaptures, calibration time and travel are never billed.

Every participant signs a written release granting commercial training rights before capture begins, and is paid a published hourly rate rather than a per-task bounty. Faces and licence plates of bystanders are blurred, audio is scrubbed of identifying speech, and any clip where a bystander cannot be de-identified is discarded rather than shipped. You receive a perpetual, irrevocable, buyer-owned commercial license, the consent artefacts for the batch, and a data card listing jurisdictions of capture.

RLDS, LeRobot, WebDataset, HDF5, zarr, and Rerun .rrd. Every episode carries int64 nanosecond timestamps per stream, camera intrinsics and extrinsics, handJoints[2][21][3], bodyJoints[24][3], instance masks, and action segments. Converters ship as readable source, not a binary, so you can retarget the schema to your own loader.

Spec conversation the same week, first validated batch in a median of nine days from a signed spec. Thin coverage areas — construction, clinical, agriculture, hospitality — take longer to staff, typically three to four weeks to first batch, because operator vetting in those environments is the bottleneck.