Egocentric Data for Robotics Research Labs

A research lab usually needs first-person human video that matches one specific question: a family of tasks, a type of environment, a label set a method depends on. That means the dataset has to be scoped by the lab, with documented capture, labels, consent and licensing, not assembled from whatever a public release happens to contain.

This site names AI research labs among the teams it serves, and its Insta360 program page describes a workplace-routine collection for an embodied AI research lab. It names no universities or academic labs as buyers, no lab or researcher as a client, and no academic pricing or grants, so this page makes none of those claims.

The data is human capture. A contributor wears the rig and performs tasks from the spec, so the output is first-person human video with synchronized streams and labels, not robot trajectories or teleoperation logs.

Use a public dataset when it already matches your tasks and you are exploring. Commission custom data when the gap is specific: a task, an environment or a label the public releases do not cover, or terms that do not fit how you intend to use the data.

The EgoVerse authors state the problem directly: existing human datasets are “often limited in scope, difficult to extend, and fragmented across institutions.” Their answer is a shared platform that takes contributions from individual researchers, academic labs and industry partners. A custom collection is the other route: you write the scope and we capture to it.

Terms matter as much as content. Many public egocentric datasets are released for academic benchmarking under terms that do not grant commercial training rights, which this site covers in its guide to research-only versus commercial dataset licenses. The first table compares EgoVerse and Egocentric-10K, using their official sources, with a custom collection. Where those sources do not state something, the table says so.

The skill spec is a written document that fixes what will be captured and how it will be judged before any footage is recorded. We draft the first version with you on a call. The second table lists what it sets and where the site states each item.

Many research programs start with a Spec Pilot, a small first batch that proves the spec and accept criteria on real footage before volume scales. The pilot is quoted for reach, timeline and price like any other order.

Anything the spec leaves open stays open in writing. Contributor counts, scene counts, hours and regions are set per program in the skill spec; this page states none of them.

The instrumented tier is a head-mounted stereo rig. It records first-person RGB, with metric depth recovered from the calibrated stereo rig rather than a standalone depth sensor, and 6-DoF IMU data at 200 Hz from head- and wrist-mounted sensors. The wider stack is ten streams: a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, instance segmentation and action labels.

Labels follow the delivered episode schema: hand_joints with hand_visible flags, body_joints, contact_events, instance_masks, objects, action_segments and gaze_target. Which of them a program attaches is set in the skill spec. The third table maps each part of the setup to where it is recorded in the episode.

Calibration is part of the data. Each session is solved twice, before and after capture, with the drift between the two written into the episode metadata, so a lab can filter on it rather than trust a summary.

Validation runs against signed accept criteria written into the spec. A reject log ships with every batch. An episode is accepted only if RMS reprojection stays under 0.28 px and calibration drift is within tolerance; otherwise it goes to the reject log, which you can see.

At ingest, episodes are machine-checked against the criteria in the fourth table, and a human reviewer then checks each episode against the task spec. Rejected items are recaptured at our cost rather than shipped with a caveat, and billing is per validated hour.

This describes our acceptance process. It is not a claim about how a delivered dataset performs in any particular model or experiment.

The license is perpetual, irrevocable and worldwide, with the data buyer-owned and non-exclusive by default. Training, evaluating, fine-tuning and deploying models, including commercial ones, is permitted. Exclusivity is available per skill line or per environment as a separate term. The executed agreement controls.

The site states no terms for publication or redistribution of results or derived data. If your lab needs them, raise it before the spec is signed so it is settled in the agreement rather than assumed.

Delivery is in the format your stack reads: RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. The HDF5 layout, including stream names such as streams/depth_mm and streams/imu_head, is on the custom HDF5 page and the episode schema page.

Every contributor signs a written release before capture, and each delivered episode links to that record through consent artefact IDs in the batch data card. The consent chain has six checkpoints, from before capture to after delivery, and a break at any one discards the footage rather than shipping it at a lower grade.

At ingest, automated detection finds faces, plates and screens and blurs them. A human reviewer confirms that no un-consented person is identifiable in any frame, and footage that cannot be de-identified is discarded.

The provenance page publishes the release and the anonymization page describes the full pass. We make no claim about institutional review board approval or any regulation; if your institution needs its own review, set the requirement per program in the skill spec. None of this is legal advice, and the executed agreement controls.

Describe the question: the tasks, environments and labels your experiment needs. That becomes a written spec, quoted for reach, timeline and price before any capture begins, and many labs start with a Spec Pilot.

Where coverage for your tasks and regions is deep, the median time from signed spec to first validated batch is a median of 48 hours; where it is thin, three to four weeks while recruiting expands. Anything this page does not state is set per program in the skill spec.