CorpusMarket Robotics
Custom data collection for robotics
The footage that is hardest for a lab to get is ordinary work in the places it really happens. CorpusMarket takes your brief and introduces you to a programme whose recordings come from the crews of real businesses. We do not sell datasets ourselves and we quote no prices.
By Linus Näslund, Co-Founder
Contents07
Custom data collection for robotics, through real businesses
Much custom data collection for robotics happens in spaces a lab or a vendor controls: a rented room, a mock kitchen, a few hired operators. That gives clean footage of one place, one set of objects and a handful of people.
What is scarce is footage from where work really happens. A commercial kitchen during service, a mechanic's bay, a job site and a stockroom all belong to businesses, and the people who know the work are their employees. CorpusMarket takes enquiries from businesses whose crews could record their own work, checks the fit, and introduces them to a recording programme run by our partner Amelia Robotics. For a lab, that programme is a route to real world data for robotics from many workplaces instead of one staged set.
What can be captured
The programme produces human demonstration data for robotics in one form: first-person video of people doing their normal work. A worker wears a phone on a hardhat or headband mount with the camera pointing down at the hands, starts a recording in the programme's app, does an agreed task and stops it. Recordings pass a quality review before they count.
The work it covers today:
- Plumbing, HVAC and electrical work, carpentry and construction
- Auto services
- Warehouses, manufacturing and food manufacturing
- Restaurants and commercial kitchens
- Commercial cleaning, hotels, retail and grocery
- Farms and nurseries, event venues and hair salons
The value is in the spread. The same task recorded across many sites, tools, lighting conditions and people is closer to what a robot will meet than the same task repeated in one room. The guide to egocentric data collection describes the format, and the industries section shows what the tasks look like in each line of work.
What to put in a brief
A brief does not have to be long. It has to be specific enough that someone can tell whether a recording meets it.
| Item | What to state |
|---|---|
| Tasks | The activities you need, with a clear start and finish for each, and what counts as a variation |
| Environments | The kinds of workplaces, and how many different sites you want the footage spread across |
| Hours | Total volume, the pace you need it at, and whether hours per task or per site matter |
| Camera perspective | Head-mounted first-person is the default here; say what must stay in frame |
| Acceptance criteria | What makes a recording usable for you, and what makes you reject one |
| Delivery format | File format, resolution and frame rate, metadata, labels or segmentation, and how you want it transferred |
If you need anything beyond what a phone camera records, such as depth, hand pose or extra camera angles, say so in the brief. Do not assume it is included.
How a collection usually runs
Custom collections in this field tend to follow four stages. The specifics of any one project are agreed between the lab and whoever records.
- Brief. The lab describes the tasks, environments, volume and criteria. Open questions are settled before anything is recorded.
- Pilot. A small batch is recorded and delivered. The lab checks it against its criteria, and the brief is tightened where the footage shows a gap.
- Collection. Recording runs at volume against the agreed brief, with review along the way.
- Delivery. Approved footage is handed over in the agreed format, in batches or at the end.
The pilot is the stage that saves the most time. Problems with framing, task boundaries or metadata are cheap to fix after ten hours and expensive after a thousand.
What CorpusMarket does
CorpusMarket is an independent broker and is not part of Amelia Robotics. On a lab enquiry we do three things: read the brief, come back within 2 business days with any questions, and introduce you to the programme if the need matches what it records.
We do not sell datasets, and we do not quote prices. Scope, price, rights and delivery are agreed between the lab and the programme. For background on how rights in footage are usually handled, see video data licensing. If you are comparing routes, our overview of robotics data collection companies sets this approach beside the alternatives.
Honest limits
This kind of data is useful for some things and wrong for others.
- It is not robot-embodiment data. First-person human video contains no joint positions, gripper states or robot actions. If your method needs demonstrations recorded in the robot's own body, you need teleoperation. The comparison of teleoperation and egocentric video goes through the trade-off.
- Real workplaces set their own boundaries. Recording is voluntary for each worker, and what can be filmed at a given business is gone through case by case before it starts. Some tasks and some sites will not be available.
- Volume follows participation. Hours depend on how many businesses and workers take part in a given line of work, so a narrow task in a rare setting takes longer to fill than a common one.
- Geography. The programme collects mainly in the United States today.
- You may not need a custom collection. If a public dataset already covers your tasks, start there. Our guide to egocentric video datasets describes what exists.
Send a brief
Use the form below. It asks for your lab or company, your country, what you need captured, and your name, email and phone. We confirm your email with a six-digit code and contact you within 2 business days.
Tell us what you need captured
No password. The code confirms your address and opens your page.
We confirm your email with a six-digit code and contact you within 2 business days.
Questions and answers
Frequently asked questions
What is custom data collection for robotics?
Recording made to a lab's own specification instead of taken from an existing dataset. The lab sets the tasks, environments, hours, camera perspective and acceptance criteria, and the footage is recorded to match.
Does CorpusMarket sell robotics datasets?
No. CorpusMarket is an independent broker. We take your brief and introduce you to the recording programme run by our partner Amelia Robotics, and scope, price and terms are agreed between you and the programme.
Can first-person human video be used to train a robot?
Labs use it, usually alongside data recorded on robots. It shows how a task is done by a person in a real setting, but it contains no robot actions, so the model has to bridge the gap between a human hand and the robot's own body.
What does a custom collection cost?
We quote no prices. Cost depends on the tasks, the environments, the hours and the acceptance criteria, and it is agreed between the lab and the programme once the brief is clear.
Where is the footage recorded?
In real workplaces, by the crews of businesses that have joined the programme. The programme collects mainly in the United States today, and other countries are answered case by case.
More pages
Read next
Robotics data marketplace: how labs source real-world data
There is no single robotics data marketplace where a lab can buy the hours it needs off a shelf. Real-world data comes through a handful of routes, each with its own strengths and limits. This page compares them for a data lead who has to choose.
Read moreRobot teleoperation data vs egocentric video
There are two ways to record real-world examples for a robot. A person can steer the robot through the task, or a person can do the task and film it from their own point of view. Each gives the model something the other cannot, which is why most labs use both.
Read moreVideo data licensing for AI training
AI developers license video from businesses the same way they license text and images. Whether your video is worth licensing depends on what it shows, who is in it and whether you hold the rights. This page separates the two routes, licensing footage you already have and recording new footage that a lab has asked for.
Read more