Request an Egocentric Data Sample

An egocentric data sample is real first-person footage, with the sensor and label data that ship beside it, supplied so a buyer can judge quality before commissioning a custom collection. The site describes it in two layers. The sample pack is the evaluation set: 40 episodes across 4 environments, 8 synchronized streams per episode, about 2.4 GB. The sample data outputs page is the per-recording view: two real 10-second clips with matching IMU rows, and the list of every file delivered per recording.

The site says the pack is drawn from the same production batches you would license, with every stream at native rate, rather than a curated highlight reel, and that it carries the same enrichment layers as a production batch. This page is for robot learning, VLA, humanoid and AI data teams who want to test first and then order a fast, on-demand custom collection.

A custom collection is billed per validated hour against a written spec, so the cheapest mistake is the one caught first. A sample answers the generic questions without a spec: does the schema load, are the labels usable, is sync as stated. The site makes the same split: the free sample pack is enough to sanity-check the loader and the label quality, and a Spec Pilot is enough to move a policy.

The site also lists what to check on any provider: measured cross-stream sync error, annotation depth including manipulation-grade 3D hand pose, licence terms, documented participant consent, failure-case coverage, and whether billing is per validated hour or per raw hour. A sample lets you test the first three directly on files you hold.

The site states the instrumented rig as 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 on one PTP clock, with sync of 1.12 ms median, 1.94 ms p99 and a 2.0 ms ceiling. The sample pack carries 8 synchronized streams per episode. The site does not list which eight, so the exact stream list for a given sample is agreed per request.

On the label side, the schema states per-frame hand joints as handJoints[2][21][3], body joints as bodyJoints[24][3], instance masks, action segments, camera intrinsics and extrinsics, and int64 nanosecond timestamps per stream. The table below shows what the site says the pack holds.

The sample data outputs page shows two real 10-second slices, a US home-baking recording and a home-cooking recording, each with the IMU rows for the same seconds. Faces are blurred and the web clips are downscaled to 1080 × 1080; the delivered files are the full 4608 × 4608 frame at the original length.

That page describes a head-mounted Insta360 X5 at 4608 × 4608 and 30 fps with IMU at about 1,000 samples per second. The rig specs above are stated separately and the site does not map one onto the other, so ask which capture setup a given sample comes from. The table lists the files per recording.

The site says the sample pack ships in every format it lists, so you can point your own loader at real episodes. Delivery formats are RLDS, LeRobot, WebDataset, HDF5, zarr and Rerun .rrd. Every format is a projection of one episode schema with four groups: meta, streams, annotations and calibration.

Converters ship as readable source rather than a binary, so you can retarget the schema to your own loader. A sample in a format or layout beyond those listed is agreed per request, not promised here.

Start with the loader. Open the sample in the format your training stack already reads and confirm that shapes, dtypes and units match the episode schema. Then check timestamps: every stream carries int64 nanosecond timestamps, so compute the offset between head, wrist, depth and IMU streams yourself and compare it with the stated median and ceiling.

Next, inspect the labels on real frames: overlay hand joints, masks and action segments on the video and look at occlusion, drift and boundary placement. The full pack includes annotations.json and a Rerun recording with every stream on one timeline, which makes this visual pass quick. Finally, read the licence terms and note which of your requirements the sample cannot answer. Those open questions become the written spec for a pilot.

They answer different questions. A sample shows existing footage already captured for a different purpose, so it tests the schema, loader fit and label quality. A Spec Pilot is a small, fast batch captured fresh against your own written spec, QA’d with the same reject log and coverage reporting as a full order, so it proves your spec rather than a generic example. A quote scopes the full volume.

In practice the order is sample, then pilot, then quote, though a team that already trusts the schema can go straight to a quote. The site says quotes are scoped per engagement and billing is per validated hour, with rejected episodes, recaptures, calibration time and travel never billed.

The site’s terms state that sample datasets are provided for evaluation only. Redistribution, resale or use of sample data in a production or commercially deployed model is not permitted unless a separate commercial licence has been signed. Treat the sample as a test fixture, not a training corpus.

A custom collection is different: the site states a perpetual, irrevocable, buyer-owned commercial licence, non-exclusive by default, with the consent artefacts for the batch and a data card listing jurisdictions of capture. A pilot is held to the same licence and billing unit as a full order.

The pack is offered in the sample section of the homepage, where the site says to add an email and the download link lands in your inbox. The site describes the pack as free, with no call or NDA required. It states no delivery time for the pack, so this page does not promise one.

If you want a sample shaped to your own task, send the activity and object list you need, and the format your loader reads. What a custom sample contains, how large it is and when it arrives is agreed per request. When the sample fits, write the spec and start a pilot: a spec conversation happens the same week, and a first validated batch arrives in a median of 48 hours where coverage is deep.