RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data

TL;DR

Reka is releasing RekaDaily-10k — 10,312 hours of unscripted first-person household video captured by paid contractors recording their own routines in real homes across the US, LatAm, Asia, and Africa, with roughly 1,670 hours in native 4K, under Apache 2.0. The dataset ships in two tiers: a raw tier (full sessions as recorded, no processing) and a processed & captioned tier (short clips with one language caption per clip written against the arc of the full session, not a single sampled frame). It’s positioned as a complement to Ego4D, EPIC-KITCHENS, and Egocentric-10K by covering unscripted domestic activity — laundry, dishwashing, groceries, cleanup, short two-handed fix-it jobs — the kinds of clips no one would upload to YouTube and no lab would commission via teleoperation. The corpus is produced by Claru, Reka’s data engine, which pays a network of 100,000+ collectors per accepted hour and gates every clip through a two-pass QC pipeline (signal + content) plus automated PII screening with human review on flagged frames.

Key claims

The corpus contains 10,312 hours of unscripted first-person household footage, with ~1,670 hours in native 4K, released ungated under Apache 2.0 covering commercial use and redistribution [§Introducing RekaDaily-10k].
What distinguishes this release from prior egocentric corpora is unscripted activity in real homes with detailed captions — collectors record their own routines rather than performing a task list, so sessions run long, hands leave frame, tasks get abandoned and resumed later, and other people interrupt [§Where The Footage Comes From, §The Egocentric Landscape].
Environment diversity scales with the number of collectors, not with hours recorded: each collector records in their own home, producing variation across kitchens, appliances, cabinet layouts, floor plans, lighting, clutter, packaging, signage languages, and weather that “is difficult to produce any other way” [§Environment Diversity].
Captions are written against the arc of a full session rather than a sampled frame — e.g. “loads the washer, leaves for twenty minutes, returns to unload it and folds the load” — so that a caption tied to a 20-minute pause between loading and unloading remains meaningful supervision [§From Submissions to Captioned Video].
The QC pipeline is calibrated against a labeled sample of past human verdicts and tuned so that borderline video goes to human review rather than to the bin: “a false rejection is the worst outcome” and no automated check can reject without clearing a confidence floor [§From Submissions to Captioned Video].
Automated PII screening reads every clip for recurring cases (face in a mirror or on a screen, name on a prescription label, mail) and routes flagged clips to a human reviewer who either anonymises the region or drops the clip; explicit bystander-consent instructions are given to collectors [§Consent and Privacy].
The raw tier ships straight from submission (no clipping, filtering, or captioning applied) — deliberately unopinionated so downstream teams can implement their own clipping and annotation schemas [§What You Can Train On It].
This is the second Reka open dataset release in as many months, following RekaCS2-10k (10,000 h of egocentric Counter-Strike 2 footage with per-frame action annotations and an open-source demo renderer) [§Get The Data].

Method

Data pipeline:

Collection. Claru sources footage from 100,000+ paid contractors globally. Collectors join a project, pass a qualification assessment, then record their own household activity on phones mounted on their heads and get paid per accepted hour. Sessions are unscripted — the collector picks the task and records it at their natural pace.
Quality control (two passes). Signal pass checks hands-visible, illumination, frozen frames, orientation, padding, length, task-match. Content pass checks POV vs staged, screen recording, time inflation. Every video also gets a perceptual fingerprint for near-duplicate detection at scale. Thresholds were fit by replaying historical submissions with existing human verdicts and moved toward reviewer decisions on disagreements. QC was introduced mid-collection so most (not all) of the released set went through it; everything going forward is gated.
PII screening. Automated pass reads every clip for face-in-mirror, face-on-screen, name-on-prescription-label, mail. Flagged clips go to a human reviewer who either anonymises the region or drops the clip.
Clipping + captioning (processed tier only). Video is cut into shorter clips, deduplicated via perceptual fingerprint, and captioned. Captions are written to describe where an activity sits in the arc of the full session — order, pauses, corrections — not just the sampled frame.

Raw tier: full sessions as collected, no processing applied. Processed tier: short clips + one caption per clip, language-supervision-ready.

The blog references a companion piece — World Model Data Pipeline — for the full processing details, and Claru at claru.ai for custom collection requests.

Results

No benchmark numbers — this is a data release, not a model release. The load-bearing claims are corpus-scale and coverage:

10,312 h total, ~1,670 h native 4K, Apache 2.0, ungated on HuggingFace [§Introducing RekaDaily-10k, §Get The Data].
Multi-region coverage: US, LatAm, Asia, Africa [tweet body].
Task coverage: laundry (pile → folded), kitchen cleanup, dishes, unloading groceries, wiping surfaces, reorganising rooms/closets/drawers, sweeping, trash, watering plants, clearing tables, unboxing, changing bulbs/batteries — deliberately the short two-handed jobs no one commissions teleop for [§Everyday Household Routines].
Two positioning contrasts named in the blog: (a) YouTube cooking videos are edited/staged/tripod-mounted and cut for the interesting parts, whereas world models need continuous unedited first-person recordings at natural speed in the natural mess; (b) teleoperation is precise but slow to produce and inherits the tidiness of the collection space, whereas real first-person recordings sit in between [§Introduction].

Why it’s interesting

RekaDaily-10k is the largest filed unscripted, first-person, household-scoped video corpus with per-clip captions — and it lands in the middle of an active wiki debate about whether human egocentric video can substitute for robot teleop data. HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining‘s controlled matched-budget head-to-head found egocentric video outperformed real-robot data for embodied pretraining, and EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data scaled the same idea to 20,854 hours to hit VLA-quality dexterous manipulation. RekaDaily-10k sits half the size of EgoScale but is broader (household + multi-region + unscripted) and open under Apache 2.0 rather than published-with-paper. If the HumanScale / EgoScale results hold at this corpus scale, this becomes the default open pretraining substrate for household-scoped VLAs and video world models — the same role Ego4D played five years ago for a smaller egocentric research community.

This also complements WorldModelGym: a decision-based fidelity benchmark for world models (WorldModelGym) as the second half of Reka’s post-Moonvalley-merger open-ecosystem play: WorldModelGym is the evaluation substrate for physical-AI world models, RekaDaily-10k is the training substrate. Together with Reka and Moonvalley Join Forces to Advance Models and Infrastructure for Physical AI they read as a deliberate positioning around household + real-world physical AI as the target application. The Claru data-engine framing (“this is what our data engine produces daily”) is closest in spirit to Egocentric-1M — largest egocentric video dataset (Build AI / Eddy Xu announcement) Egocentric-1M and Computer Use Large: 48k screen recordings (~12,300 hours) of professional software Computer Use Large — third-party data-engine companies rather than model labs — but at Apache-2.0 rather than gated release.