03Buyers
Key facts
- Three data types: residential egocentric, commercial egocentric, and gaming with synchronized inputs.site data section; Data Collector Agreement §1.2
- Every identifiable person in footage signed a release, and every commercial site gave written permission.Privacy Policy §2, "If you appear in footage"
- Footage reaches buyers only after it passes quality review.Privacy Policy §4
- Datasets never include collectors' names, contact details or payment data, and buyers are contractually barred from identifying anyone in the footage.Privacy Policy §4
- TaskCapture works with hundreds of data collectors in 50+ countries.homepage network block
What egocentric video is, and why robot teams use it
Egocentric video is shot from the point of view of the person doing the task, usually with a head-mounted camera. You see hands, tools and objects from the distance and angle the work happens at, and the steps come in the order the person did them. For the basics, read our explainer on what egocentric video is.
A robot's own camera sees much the same view while it manipulates objects, so robot-learning researchers use this footage for pre-training. R3M pre-trained a visual representation on Ego4D human video and reported that it helped robots learn manipulation tasks from few demonstrations. What teams want from it is data efficiency: a representation learned from varied human video can let a robot pick up a new task with fewer of its own demonstrations. Our guide on how human video helps robots learn covers that research and where it stops.
Three data types
Each request can draw on one type or several.
| Residential | Commercial | Gaming | |
|---|---|---|---|
| Settings | Collectors' own homes | Restaurant prep lines, warehouses, retail, workshops and trades, hotels, farms | The collector's own gaming setup |
| Typical tasks | Cooking, dishes, laundry, cleaning | The collector's normal shift work | A play session |
| What's in frame | The collector's hands and the task, from a head-mounted camera | The collector's hands and the task, from a head-mounted camera, on site with written permission | The game screen (no camera) |
| Extra signals | Recording metadata | Recording metadata | Every keyboard, mouse or controller input, in sync with the screen. No voice, chat, gamer handles or game telemetry |
From request to footage
We don't sell a fixed catalogue. TaskCapture collects to your specification.
- 01
You tell us what you need
Tasks, settings, hours and deadline.
- 02
We match collectors
People who already do those tasks, in the places you need.
- 03
They record the work
For physical work, a small head-mounted camera worn during normal tasks; for gaming, the screen with every input.
- 04
You get the footage
Video and metadata in the format you ask for, after quality review.
A worked example: scoping a commercial-kitchen request
The team below is invented, and the example uses no TaskCapture figures.
Say a robotics team is pre-training a manipulation policy for commercial food prep. Its request names the data type (commercial) and the tasks ("chopping, portioning and plating on a prep line"). It adds a target number of hours, a deadline and the countries that matter, then the format: video plus timestamps and device metadata, at a resolution and frame rate the team states.
We match collectors who already work prep lines in those countries. Each one gets written permission from the site and a release from every coworker who appears on camera. Footage passes quality review before delivery, and the team gets it in the format it asked for. Nothing in the request names the collectors, and nothing in the delivery does either: datasets never include their names, contact details or payment data.
Consent, releases and provenance
Every identifiable person on camera has signed a release. Collectors must get a signed release from every identifiable person they record and, in any commercial setting, written permission from the site's owner or manager. A person under 18 appears only with a release signed by a parent or guardian; otherwise the footage is rejected and deleted.
TaskCapture does not run face or voice recognition and does not create biometric identifiers. Collectors agree to each project before they record, and footage from data suppliers follows the same consent rules as footage from TaskCapture's own collectors. Buyers' contracts limit use to developing, training, evaluating and improving AI and machine-learning models, and bar them from identifying anyone who appears. The full rules are in our Privacy Policy.
Check provenance on any dataset, public or custom. When the Data Provenance Initiative audited widely used text datasets, it found licence omissions above 70% and error rates above 50% on popular dataset hosting sites. Public egocentric datasets come with their own terms as well. EPIC-KITCHENS, for example, is published under a non-commercial licence. Our guide to training-data licensing and consent lists the questions to put to any supplier.
Metadata and formats
Recordings carry metadata such as the device used, timestamps, and location or GPS data where the recording embeds it. Gaming sessions add the input log: every keyboard, mouse or controller event, in sync with the screen.
We deliver video and metadata in the format you ask for. Resolution, frame rate, minimum volumes and lead times are agreed per request: tell us yours in the form. Our egocentric dataset spec checklist lists the fields teams usually specify.
Custom collection compared with public datasets
Public egocentric datasets are a common research baseline. The Ego4D paper reports 3,670 hours of daily-life video from 931 camera wearers in 74 locations across 9 countries. EPIC-KITCHENS-100 has 100 hours of unscripted kitchen activity in 45 environments, recorded on head-mounted cameras, with 90,000 annotated actions.
| Public research datasets | Custom collection (TaskCapture) | |
|---|---|---|
| Tasks and settings | Fixed: what was recorded | Your tasks and settings, recorded on request |
| Volume | Fixed | Agreed per request |
| Licence | Set by each dataset; some are non-commercial | Agreed in your contract |
| Consent | Set by each project's protocol | Signed releases; written site permission |
| Best for | Benchmarks, baselines, research pre-training | Coverage of the tasks and places your model will meet |
Plenty of projects use both: a public set for baselines and custom footage for the tasks the model will meet. Our guide to the main public egocentric datasets compares the public options, and custom versus off-the-shelf data helps you choose between them.
How to judge egocentric data before you buy
Ask for a sample and documentation before you commit. Datasheets for Datasets lists the questions a dataset's documentation should answer: why it was created, what it contains, how it was collected and what it should not be used for. With egocentric video, also check that hands and the task stay in frame, that each task runs start to finish, and that the metadata matches the video. Our guide to evaluating video dataset quality turns those checks into a checklist.
Common mistakes in data requests
- Asking for "kitchen video". A category is hard to scope; "chopping and plating on a prep line" names a task we can match collectors to.
- Leaving the format for later. State video format, resolution, frame rate and metadata at the start, since they are agreed per request.
- Leaving out where the robot will work. If it will run in a particular kind of site or country, say so.
- Treating consent as a box to tick. Ask how the releases and site permissions were collected as well as whether they exist.
- Skipping the sample. A few clips show framing and task coverage faster than a spec sheet does.
What to put in a request
We can scope a request quickly when it covers these points:
- The data type: residential, commercial, gaming, or a mix.
- The activities and places, as specific as you can make them (for example, "restaurant prep: chopping and plating").
- How many hours of footage and how many tasks you need, and your deadline.
- Video format, resolution, frame rate and the metadata you need.
- Countries or regions, if they matter for your model.
Questions
Is this an off-the-shelf dataset or custom collection?
TaskCapture collects to order. You describe the tasks, settings, hours and deadline, and we match collectors who already do that work in the places you need.
How is consent handled?
Every identifiable person on camera signs a release, commercial sites give written permission, and people under 18 appear only with a parent's or guardian's release. TaskCapture runs no face or voice recognition, and buyers are contractually barred from identifying anyone in the footage.
Can you collect in a specific country or setting?
Tell us the countries and settings you need in your request. TaskCapture works with hundreds of data collectors in 50+ countries, in both commercial and residential settings, and we confirm what we can cover when we scope your request.
What formats do you deliver?
Video and metadata in the format you ask for. Resolution, frame rate, minimum volumes and lead times are agreed per request.
How long does a collection take?
Lead times are agreed per request. Put your deadline in the form, and we confirm what we can cover, and by when, when we scope your request.
What can we use the data for?
Contracts limit use to developing, training, evaluating and improving AI and machine-learning models, and bar buyers from identifying anyone in the footage.
Does gaming data include a camera or voice?
No. Gaming data is the game screen with every keyboard, mouse or controller input in sync. There is no camera, and no voice, chat, gamer handles or game telemetry.
How does this compare with teleoperation data?
Teleoperation data records a person driving a robot through a task. Egocentric video records people doing the task with their own hands. Our comparison guide explains when each one helps.