CLARU
Data Collection

Household Egocentric Video for Humanoid Robots

344K+ hoursHousehold egocentric video accepted for humanoid training
summary.md

Challenge:Humanoid robots need to learn household work from human perspective, at human height, in the messy environments homes actually are.

Solution:We embedded capture with a household contributor program running in 32 countries.

Result:The program delivered the largest commercially licensable household egocentric corpus we know of: 126 tasks already exceed 100 accepted hours each and 62 exceed 1,000 hours each, giving humanoid and world-model teams both breadth (220+ tasks, 32 countries) and depth where it matters.

344K+ hoursAccepted household egocentric hours
220+Distinct household tasks captured
32 countriesCountries with accepted household capture
31,000+Approved household contributors
// THE CHALLENGE

Humanoid robots need to learn household work from human perspective, at human height, in the messy environments homes actually are. Lab demonstrations are staged and short; open datasets offer neither commercial licensing nor the task breadth general-purpose home robots require. The buyer needed hundreds of thousands of hours of consented, commercially licensable first-person video spanning the full range of household activity — cooking, cleaning, laundry, dishes, home maintenance — captured on real cameras in real homes, delivered in continuous weekly waves.

// OUR APPROACH

We embedded capture with a household contributor program running in 32 countries. Contributors wear chest- or waist-mounted cameras that approximate humanoid camera height, then perform their normal routines against a 220+ task taxonomy that covers the activities a home robot actually performs: meal preparation, kitchen cleanup, laundry, cleaning and tidying, home maintenance, and more. Collections are quota-driven — every task bucket has explicit targets — so delivery stays broad rather than clustering on easy tasks.

Every submitted clip passes an acceptance gate before hours count: task match, camera stability, and coverage of the full activity are reviewed, and rejected submissions never enter the delivered corpus. Accepted hours are tracked per task, per country, and per week, with weekly delivery waves: the program peaked at 52,036 accepted hours in a single week.

The taxonomy is versioned so new task categories can be added mid-program without breaking delivery, and per-country quotas keep the corpus from collapsing onto a single region's homes, kitchens, and habits.

  1. 01RecruitOnboard household contributors across target countries with consented capture agreements
  2. 02CaptureHands-free chest/waist-mounted recording at humanoid camera height, driven by task quotas
  3. 03ReviewAcceptance gates on task match, stability, and coverage — rejected clips never count
  4. 04DeliverWeekly delivery waves with per-task, per-country tracking; peak week 52,000+ accepted hours
// FOOTAGE

Household footage on Explore

Public household clips from Claru Explore, shown as muted previews. Explore draws on several capture programs, so mounting and framing vary from clip to clip.

Placing two cups into a kitchen cabinet
Sorting and folding laundry, overhead view
Loading a top-load washing machine
Grating food onto a plate at the dining table
Browse Kitchens on Explore
// QUALITY GATES
  1. Task matchThe clip shows the task it was submitted against in the taxonomy.
  2. Camera stabilityFootage is steady enough to follow hands and objects through the activity.
  3. Full-activity coverageThe clip covers the activity from start to finish, not a fragment of it.
  4. Rejected clips never countOnly accepted clips add hours toward task, country and weekly quotas.
// RESULTS
344K+ hoursAccepted household egocentric hours
220+Distinct household tasks captured
32 countriesCountries with accepted household capture
31,000+Approved household contributors

Task depth in the accepted corpus

Distinct household tasks captured
220+
Tasks with 100+ accepted hours each
126
Tasks with 1,000+ accepted hours each
62

Bar length is proportional to the number of tasks. Hours are accepted hours only.

// WHAT THE CUSTOMER RECEIVED

Accepted first-person household clips, delivered in weekly waves and tracked against the task taxonomy.

First-person video
Hands-free household footage recorded at humanoid camera height in contributors' own homes.
Task
Every accepted clip maps to a task in the versioned 220+ task taxonomy.
Country
Tracked per clip, so quotas can be balanced across 32 countries.
Weekly waves
Accepted hours reported per task, per country and per week; the peak week passed 52,000 accepted hours.
Licensing
Consented capture, delivered under commercial training-use licensing.
// IMPACT

The program delivered the largest commercially licensable household egocentric corpus we know of: 126 tasks already exceed 100 accepted hours each and 62 exceed 1,000 hours each, giving humanoid and world-model teams both breadth (220+ tasks, 32 countries) and depth where it matters. Continuous weekly waves let the buyer steer quotas toward underrepresented tasks as training runs surface gaps.

// WHY IT MATTERS

Household assistance is the flagship use case for humanoid robots, and it is the hardest environment to source data for: it is private, unstructured, and culturally varied. A consented, quota-driven household program — rather than scraped or staged footage — is what makes this corpus usable for commercial training, and the 32-country spread is what makes the resulting policies generalize beyond one region's homes. Reka has also released part of the network's household footage openly as RekaDaily-10k, so teams can inspect the footage type before scoping a program.

// DECISIONS WE MADE
Mount cameras at chest or waist heightwhyIt approximates where a humanoid's head cameras sit and keeps both hands in view during manipulation.
Run on quotas, not open submissionwhyEvery task bucket has a target, so delivery stays broad instead of clustering on easy tasks.
Version the taxonomywhyNew task categories can open mid-program without breaking delivery.
Set quotas per countrywhyKeeps the corpus from collapsing onto one region's homes, kitchens and habits.
// KEEP EXPLORING
// RELATED
// SOURCES
  1. [1]RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data (Reka)Reka's open release drawn from the household portion of Claru's collector network. A separate release, not this program's delivery.
  2. [2]RekaDaily-10k raw tier on Hugging FaceApache 2.0, ungated.
// FAQ

Ready to build your next dataset?

Tell us about your project and we will scope a plan within 48 hours.