Egocentric

First-person capture from a head-mounted rig: skilled trades at work in operating businesses, and everyday manipulation at home.

Behaviour cloning and VLA training where the policy has to see what the hands see: the grasp, the tool, and the order a task is really done in.

The shelf behind these runs to roughly 15,000 episodes and 1,200 hours, shot across 70+ operating businesses. Counted off our own collection, rounded down.

135
Delivery files: 997 min
Playable here: 16
Skill groups: 111
Distinct tasks: 18
Industries: 29
Workplaces: 65
Operators

Trim the neckline of a customer's haircut · 1/135

These are web previews, not the data. Each clip is downscaled to play in a browser. The delivery files are full resolution and carry every stream the rig recorded.

Camera configurations

Four rigs, one delivery pipeline.

Every configuration covers the same spread of trades, workplaces and difficulty — what changes is how much geometry ships with the picture. Off the shelf or collected to your spec, and convertible to any format you work in: MCAP, LeRobot, RLDS, HDF5.

Mono egocentric (1 camera)One camera on the head — smartphone or GoPro — with a scale reference in frame. No depth and no parallax, and in exchange the widest spread of trades and workplaces and the fastest ramp we run.7 playable · 1.7 h

Stereo egocentric (2 cameras)Two synchronised eyes and a 200-400 Hz IMU on one clock — true depth, not a single RGB feed. Enough parallax for VIO and hand pose, and the rig most of the published catalogue is shot on.107 playable · 13.1 h

Advanced stereo egocentric (6 cameras)Six cameras in three stereo pairs on a hardware-synced clock — frame-level, no post-hoc alignment. 100° of coverage keeps both hands in frame, so a grasp hidden from one pair is still visible in another, with per-lens calibration and a 200-400 Hz IMU on the same timeline.17 playable · 0.8 h

Egocentric with wrist cameraThe head view plus a camera on each wrist, time-synced to it. Keeps the grasp in frame at the moment the head camera loses it behind the object.3 playable · 0.4 h

What we run in this category

Workplaces: 29
Distinct tasks: 111
Operator jobs: 33
Graded hard: 26%

Measured across the samples published here, not the full shelf.

Delivered as LeRobot v3 · MCAP · HDF5 · RLDS · mocap to FBX / BVH / SMPL

Advanced annotation

The video is the smallest part of the delivery.

Open the explorer

Head movement: 6-DoF head pose per frame, from the rig's own IMU and VIO rather than estimated after the fact.
Hand pose: Per-finger joint positions on the operator's hands, tracked through the grasp.
Episode structure: The session cut into episodes and steps, so a policy trains on a task rather than on a video.
Task and environment: What was being done, where, and how hard it was graded — carried on every frame.

Explorer loads as you reach it

A real recording with its tracks on it — scrub the timeline to see the pose follow the footage, or switch sample and segment at the top. Every delivery in this category can ship the same tracks.

Off the shelf · mono phone · household and daily

A different shelf from the 1,200-hour stereo one above, not a subset of it: one phone on the head, everyday manipulation at home, licensed as a pack rather than collected to a brief. Both are called off the shelf and their totals are counted separately — the figures below are this corpus only. Every clip in the five sampled skills plays here, sixty-one of them.

12,900+
Episodes: ~72 h
Total footage: 9
Skill domains: 10-30 s
Per clip: 30 fps
Frame rate: Household · daily
Setting: 7 business days
Lead time

Up to 4K, shot in portrait. ~160-315 GB by resolution tier. Episode counts are the catalogue's own estimates, from each skill's share of the corpus.

Pick up a cup

15 playable · 49% of the corpus · ~6,321 episodes

.mp4

Use a screwdriver

15 playable · 13% of the corpus · ~1,677 episodes

.mp4

Pick up an item from a refrigerator

10 playable · 8% of the corpus · ~1,032 episodes

.mp4

Wipe a surface

6 playable · 5% of the corpus · ~645 episodes

.mp4

Cut with scissors

15 playable · 1% of the corpus · ~129 episodes

.mp4

Four more skills are in the pack with no sample folder published: open and close a door, use a hammer, arrange utensils on a rack, fold and stack cloth. Ask and we will send clips from any of them.

What this category is made of

Measured across the samples published on this page, not the shelf behind them — these are the hours you can press play on.

Skill group16

Clothing & Laundry3.1 h · 22
Tool Use & Technical Manipulation2.1 h · 16
Cleaning & Sanitation2.0 h · 15
Pick and Place / Object Handling1.5 h · 13
Packing & Bagging1.4 h · 11
Assembly & Installation1.3 h · 9

+10 more

Settingscale

Non-residential14.8 h · 122
Residential1.9 h · 13

Difficultyscale

easy4.5 h · 44
medium5.0 h · 35
hard3.6 h · 28

Trade33

Tailor / seamstress2.3 h · 16
Cook1.9 h · 15
Cleaner1.7 h · 14
Mechanic57 m · 9
Factory worker42 m · 5
Florist41 m · 4

+27 more

Industry18

Clothing and Fashion4.0 h · 26
Industrial Manufacturing1.5 h · 14
Automotive and Transport1.4 h · 13
Creative Workshops1.2 h · 8
Backyard / Yard1.2 h · 8
Repair Services1.1 h · 8

+12 more

Site type31

Household2.7 h · 11
Backyard1.2 h · 8
Handcraft Workshop1.0 h · 7
Sewing Factory1.0 h · 9
Workshop50 m · 17
Construction Site46 m · 7

+25 more

Two ways to buy

Off the shelf

Already recorded, indexed and cleared. Play any of it on this page, read the telemetry beside it, then licence the sets you want. No collection lead time — the files exist.

1,200 h of stereo capture, 100% human QC-ed
15,000 episodes averaging 4m 49s
70+ operating businesses · 35 location types · 100 operator professions
3,000 task types across 17 skill groups · 82% graded medium or hard
135 samples playable on this page

Collected to your spec

You name the environment, the device, the task mix and the volume. We design the capture, run it in operating businesses with paid professionals, and hand back the same delivery shape as the shelf.

6 rig configurations
First delivery in 5-10 business days
Up to 20,000 h / month

Either route delivers as LeRobot v3 · MCAP · HDF5 · RLDS · mocap to FBX / BVH / SMPL

Licence

Commercial use: Licensed for commercial use
Consent: Recorded per session, on every episode
Where it was filmed: Operating businesses only. No private residences, no staged studios.
Scope: Whole corpus as one pack, or single-skill licensing on request
Exclusivity: Non-exclusive by default. Exclusive terms on request.
Model training: Permitted, including derived model weights
Redistribution: Not permitted. Licensed to the buyer, not for resale.
Restricted subsets: Some manufacturing-environment data is under NDA and is not published here

Indicative terms. The licence agreement governs.

Delivered before, on these terms

Egocentric Data for a Robot Foundation Model

Real-World Video to Ground a World Model

Teleop Cold-Start for a Manipulation Startup

100%: of the capture fleet in calibration
100%: of episodes human QC-ed
70+: operating businesses recorded in
82%: of episodes rated medium or hard

Fleet figures describe the off-the-shelf corpus, the line with published numbers.

Operators stay anonymous

Operators appear only as an opaque id. Job, experience band and handedness ship with the record because a policy needs them; names do not.

Two ways into the full set

Previews on this page are open. Full resolution files, every camera, and the accompanying telemetry sit behind one of these.

Request the sample pack

Tell us the task, environment and modality you are training for. We send the matching sample files and the schema documentation.

Reply within one business day

Request access

Enter a passcode

If we have already spoken, your passcode opens the full sample set and the download links without another form.

Seven day session

Enter passcode