Open data release · October 2026
RekaDaily-10k: 10,000+ hours of real household video, collected by the Claru network
Reka has released RekaDaily-10k, an open dataset of unscripted first-person footage from inside real homes. Every hour of it was recorded by paid collectors in the Claru network. Here is how that footage was gathered, what is in the release, and where to get it.
- Raw tier
- 10,865 hours
- Processed + captioned
- 10,200 hours
- Captioned clips
- 6,373,064 clips
- With IMU (has_imu)
- ~3,000 hours
- Native 4K
- ~1,670 hours
- Ungated on Hugging Face
- Apache 2.0
Footage nobody uploads
Models that act in the physical world, from household robots to world models, learn from long, continuous first-person views of people actually doing things. The internet has almost none of it. Online video is edited, staged and cut down to the interesting parts. A full, unedited load of laundry is not something anyone posts.
So this kind of footage has to be commissioned, and that means paying people to record their own lives with a camera on their head. That is what the Claru network does. RekaDaily-10k is the household slice of it, released openly so research teams can build on it.
How the Claru network collected it
Claru is a paid collection network of 100,000+ collectors recording the physical world. The footage in RekaDaily-10k comes from a subset of them, recording in their own homes (and some workplaces) on head-mounted and handheld phones, across several regions.
- 01
Join a project
Collectors apply to a household capture project and pass a qualification assessment before recording anything.
- 02
Record their own routine
A phone in a head mount records real chores at the pace they actually happen. No task list, no staging.
- 03
Submit and get checked
Submissions run through automated quality control, and borderline video goes to human review rather than the bin.
- 04
Get paid per accepted hour
Collectors are paid contractors. Pay follows accepted footage, which keeps the incentive on usable video.
Every collector adds a new home
Because each person records where they live, the number of distinct environments grows with the number of collectors, not the number of hours. That brings different kitchens, appliances, floor plans, lighting, clutter, packaging and languages on the labels. It is very hard to get that range from a lab or a handful of staged sets.
Unscripted on purpose
Collectors record real chores rather than act out a list. Sessions run long, hands leave the frame, tasks get dropped and picked up again, and someone walks in halfway through. That mess is what a robot will meet in a real home, and it is the part that is hardest to arrange deliberately.
What household capture looks like
Kitchen and laundry footage recorded by Claru collectors, from the public samples on Explore. Washing up, sorting, ironing and loading the washer: the short, two-handed jobs that are on everybody's list and that nobody would set up a teleoperation session for.
5-second previews · muted · more on /explore/egocentric
From submissions to captioned clips
The release comes in two tiers. The raw tier is footage as recorded, so teams can apply their own clipping, filtering and labels. The processed tier has been through quality control, cut into shorter clips, deduplicated and captioned.
- 412,050 videos
- 80 TB
- ~3,000 hours with IMU
- quality control
- clip
- deduplicate
- caption
- 6,373,064 clips
- 74.2 TB
- caption per clip
Quality control checks every submission for the failures that actually recur: hands never in view, lighting too dark to use, frozen or duplicated frames, the wrong orientation, clips too short, and footage that does not match the assigned activity. A second pass looks for video that is not really first-person, is staged or screen-recorded, or has been padded to stretch paid time. Thresholds were tuned against past human review decisions, and a borderline clip goes to a reviewer instead of being rejected automatically. Quality control was introduced partway through collection, so it covers most but not all of this release; everything recorded now passes through it.
Captions describe where an activity sits in the whole session, not just one sampled frame, which matters when someone loads the washer, leaves for twenty minutes and comes back to unload it. Reka's data team covers the wider processing stack in World Model Data Pipeline.
What's in the release
- Captioned clips.Each processed clip carries a text caption plus resolution, frame rate, duration, a motion class, a quality tier, any on-screen text, example question-and-answer pairs, and the collector's own activity labels with a timeline for the source video.
- IMU on about 3,000 hours. Sessions recorded with the phone's motion sensors ship acceleration, rotation and per-frame timestamps on the same clock as the video, for frame-accurate joins.
- Native 4K on roughly 1,670 hours. A higher resolution than most large egocentric corpora.
- Apache 2.0, ungated. Commercial use and redistribution are allowed with attribution. It follows Reka's earlier open release, RekaCS2-10k.
Field-level documentation, loaders and the shard layout are on the Hugging Face dataset card and in Reka's announcement.
Consent and privacy
This is footage from inside people's homes, so how it was gathered matters as much as how much there is. Nobody in the network records by accident: collectors sign up as paid contractors and choose to wear the camera for each session. Other adults in the home have to agree before they are filmed, and anyone who hasn't is meant to stay out of shot. Clips that still show an identifiable person who isn't taking part are held back for review. Reka's announcement sets out the full consent and privacy process.
Before upload, every video had its container metadata removed: GPS coordinates, device identifiers and capture timestamps. Automated PII screening runs on top. Screening is never perfect, so Reka takes takedown requests at [email protected].
Frequently asked questions
What is RekaDaily-10k?
RekaDaily-10k is an open egocentric video dataset of unscripted household activity, released by Reka under Apache 2.0 and collected through the Claru network. It ships as a raw tier (10,865 hours) and a processed, captioned tier (10,200 hours across 6,373,064 clips).
Who recorded the footage?
Paid collectors in the Claru network, which counts 100,000+ collectors. Collectors join a project, pass a qualification assessment, record on phones in head mounts, and are paid per accepted hour. Wearing the camera is always the collector's own choice.
Can I use RekaDaily-10k commercially?
Yes. The dataset is ungated on Hugging Face under Apache 2.0, which permits commercial use and redistribution with attribution per the license terms.
Does the dataset include sensor data?
About 3,000 hours of the raw tier ship with IMU data (acceleration, rotation and frame timing, synced to the video), flagged with has_imu in the metadata. Roughly 1,670 hours are native 4K.
What if I need data the release does not cover?
Claru runs custom collection programmes through the same network and pipeline: a specific setting, set of activities, region, sensor or annotation schema. Tell us what you need and we will scope it.
Related
Need footage the release doesn't cover?
The same network and pipeline run custom programmes: a particular setting, set of activities, region, sensor or annotation schema. Tell us what your model needs and we will scope the collection.