Hands stay in frame
Contributors wear the phone on a headmount with the rear ultra-wide lens angled down at the work, so hands, tools and objects stay in view.
57 of 57audited task samples show hands at the midpoint
Egocentric data for robotics and physical AI
101,300+ hours of head-mounted, first-person video of people doing everyday household tasks, for teams training robot policies, world models and vision-language-action models.
Off-the-shelf price
US$2.40
per hour of footage
Non-exclusive commercial license
Approved inventory as of November 2026.
01 — The data
Robot policies and world models need to see hands meet objects thousands of times, in places that look like the real world. That is what this corpus is made of.
Contributors wear the phone on a headmount with the rear ultra-wide lens angled down at the work, so hands, tools and objects stay in view.
57 of 57audited task samples show hands at the midpoint
Footage comes from real residences, recorded by thousands of different people, with the clutter, lighting and layouts a lab set can’t reproduce.
5,500+contributors, Philippines-led, across 7+ countries
Clips are continuous task episodes rather than trimmed snippets, so a model sees the full sequence from setup to cleanup.
~8 minaverage clip length
Every clip is approved in human review and labeled against a 57-task taxonomy. Releases are validated against each contributor’s signed agreement.
Human QCpartial, slow and test footage excluded
02 — Coverage
Approved hours by activity family. The catalog workbook breaks this down by exact task, clip count, contributor count and country.
| Activity family | Example tasks | Relative hours | Hours |
|---|---|---|---|
| Dishes & Kitchen Cleanup7 tasks | Hand-washing dishes · loading the dishwasher · wiping counters and stovetop | 26,900 | |
| Laundry7 tasks | Folding clothes and linens · loading the washer · ironing | 22,870 | |
| Cooking & Food Prep9 tasks | Washing and chopping produce · stovetop cooking · cooking rice | 20,560 | |
| House Cleaning10 tasks | Sweeping and mopping · vacuuming · cleaning bathrooms | 19,250 | |
| Tidying & Organizing4 tasks | Picking up clutter · making beds · taking out trash | 3,120 | |
| Outdoor & Yard4 tasks | Watering the garden · sweeping walkways · bagging leaves | 2,600 | |
| Plant & Pet Care4 tasks | Tending indoor plants · filling pet bowls · grooming pets | 1,880 | |
| Groceries & Pantry3 tasks | Unpacking groceries · stocking the fridge and pantry | 790 | |
| Car Cleaning9 tasks | Washing the exterior · vacuuming seats and mats · cleaning windows | 660 |
03 — Capture spec
Contributors record in the HMD Field Recorder app on a phone worn on a headmount. The app verifies rear ultra-wide capture at 1080p and 30 fps, and every delivered clip comes with a row of machine-readable metadata.
04 — Pricing
Off-the-shelf footage costs the same for every team, so you can size a budget before the first call. Custom work is quoted per program.
Off the shelf
US$2.40/ hour
Footage from the approved corpus, licensed today.
Commissioned collection
Quoted per program
New footage collected to your brief.
Performance-based
Pay if it helps
For qualified robotics and physical AI labs.
Prices in US dollars. Exclusivity, derivative-data rights, delivery format and final legal terms are agreed in the license.
05 — Custom collection
The network and pipeline that built the corpus also run buyer-defined programs. We recruit, onboard, assign, review and pay contributors through our own recorder app, then release to the same delivery standard.
Demonstrated throughput
~5,200 approved hours
in seven days · July 23–29, 2026
Task brief, taxonomy, geography and device policy
App-verified camera, orientation, resolution and audio
Multipart upload, metadata capture and source reconciliation
Human QC, task correction, and fraud and speed controls
Rights validation, versioned manifests and controlled delivery
06 — Provenance
Licensing human data means being able to say how it was collected. Our standards are public, written into contributor agreements and enforced in review.
Contributors are 18 or older and record under a written contributor agreement. Releases are checked against it.
Other people may appear only with the permission our consent flow requires. Recording stops when someone objects.
IDs, financial documents, passwords and private screens are excluded by default. Contributors keep their own faces out of frame.
Country of collection comes from the signed agreement, never from addresses or IP geolocation.
Recordings disapproved in review are permanently removed within 30 days.
Accepted footage is paid at the rates shown in the app or assignment brief.
07 — Questions
One hour of approved footage from the existing corpus under a non-exclusive commercial license, delivered with per-clip metadata and checksummed manifests. You choose the mix by activity family, task and country. Commissioned collection, exclusivity and custom terms are quoted separately.
Yes. The public PDF catalog is free to download. Qualified teams also get the detailed workbook, with exact hours, clips and contributors by task and country, and a gated sample room of 171 preview clips covering all 57 tasks.
Through controlled cloud transfer with time-limited signed access. Every delivery ships with versioned JSON and CSV manifests listing a SHA-256 hash and byte size for each file, so you can verify everything on arrival.
Adult contributors who join through the HMD Field Recorder app, sign a contributor agreement and record assigned household tasks in real residences. They are paid for accepted footage.
The standard license is non-exclusive. Exclusivity, derivative-data rights and delivery format can be negotiated, and commissioned programs are scoped to your requirements.
Clips are recorded with device-microphone audio, encoded as AAC. If your use case needs video only, say so when you scope the order.
Request the catalog workbook and gated access to 171 preview clips covering all 57 tasks, or email us with the tasks, hours and countries you need.