Fair questions.
Straight answers on files, timing, licence, consent, price and the free sample.
What is kollected?
kollected sells egocentric (first-person) video of real manual work.
Workers wear head or chest cameras on their own shifts at real workplaces. Each clip ships as the camera's original file, with task labels from a catalogue of 354 tasks and, where the camera recorded it, the camera's own IMU on the video clock.
What is egocentric video?
Video from the worker's own point of view.
A camera on the head or chest films it, and it sees hands and tools from about where a robot's head camera sits.
What comes with each clip?
Four files, or three where the camera recorded no motion track.
The camera original (.MP4); _imu.csv, the camera's six-axis IMU on the video clock; _labels.json, the ordered task segments with their codes; and _meta.json, with workplace, lens and consent.
Is the motion data in sync?
Yes, where recorded: the camera writes its own motion data into the video file.
Its accelerometer and gyroscope are read out in microseconds on the video's clock, first frame at zero, at a nominal 200 Hz; each clip's measured rate is in its files.
Which workplaces and tasks?
16 workplaces, 70 task groups and 354 tasks.
From commercial kitchens and carpentry to welding, HVAC, construction, farms and warehouses. Every code is public, as CSV and JSON.
Catalogue →Taxonomy (CSV) →Taxonomy (JSON) →
Tasks not in the catalogue?
Ask in your quote, or from a catalogue search that finds nothing.
If we can film a task, it gets a new code; existing codes never change.
How much footage is there?
100k+ hours of first-person footage of real work.
Your quotation states the hours for your round.
Where do you deliver?
To teams in the markets listed on About, and elsewhere on request.
Files arrive by signed link, in your own bucket (S3, GCS or R2), on a shared drive or on disk.
How is it priced?
By quote.
Send the tasks, the hours and where your model fails, and you get a price and delivery dates per workplace in two working days. Only footage an independent expert confirms meets your criteria is billed.
Is there a free sample?
Yes, for teams evaluating a purchase.
A few clips from a workplace you pick, laid out like a paid delivery, sent by signed link once approved. We reply within one working day.
Free data for research labs?
Yes.
Academic groups can apply for up to 100 hours, filmed to their brief at no charge, for a citation.
How do workers consent?
Each clip is filmed by the worker wearing the camera, under a written agreement.
It covers capture, review and licensing for machine-learning training. Anyone regularly in shot signs a consent form before capture. A clip with an identifiable third-party face in more than 5% of its frames is re-trimmed, and any identifiable minor fails it outright.
What licence comes with it?
Footage, motion data, labels and metadata are supplied under the licence in your quotation.
It covers machine-learning research and development. Exclusive collection and first-look periods are available.
Which training formats?
The camera originals always ship.
LeRobot v3, RLDS, HDF5 or MCAP come beside them, on request.
Free sample →What a lab receives →
What else can you add?
In the quote: hand keypoints, language descriptions, camera calibration and narration.
The keypoints are estimated on every frame, the descriptions run from goal to step, and narration comes where the workplace has it. On request: wrist cameras, multi-view rigs, depth, tactile gloves, mocap and teleoperation demos.
Get a quote →Depth, measured →
Which cameras?
Head or chest action cameras, worn by the worker.
The site's clips come from the GoPro HERO10 Black and HERO12 Black. You set the resolution in your quote: 4K60 preferred, 1080p or any. Every delivery keeps the camera's original file.