Selling robot and first-person video data for AI training

Yes: robotics companies pay for first-person video of people's hands at work and for teleoperation demonstrations, and they spend over $100 million a year on it, by micro1's own estimate. But most of that money buys footage recorded to their specification, not archives. The people who record it are paid a median of $14 per hour of video, vendors who record to order charge $38 per hour of video, and footage sold off the shelf asks $8.88 per hour of video. Bulk first-person footage is cheap: one Indian collection startup prices it at around $1 to $10 an hour, by task complexity. What still commands a price is data a buyer cannot make cheaply: teleoperation of real robots, footage recorded to a specification with the sensors and labels a model needs, and rare settings such as factories, kitchens and fleets that outsiders cannot film. Our valuation prices unsorted first-person footage at $1 to $10 an hour, footage captured to a specification at $15 to $40 an hour and teleoperation demonstrations at $28 to $60 an hour. Because face geometry and voiceprints taken from footage can be biometric data, we take it only where the people in it agreed in writing to AI training and sale, or where nobody can be identified.

This guide is for a company that holds or could record footage of physical work (a factory, a warehouse, a restaurant, a fleet), a robotics team with teleoperation logs, and anyone weighing an offer to record. It covers what robot builders buy and why, the prices on record by kind, why bulk footage collapsed and what still sells, who pays, existing archives against recording to order, consent and faces, what makes a dataset worth more, and how a sale through AIData.Broker works. Every price comes from the page or filing its record names, read on the day shown there.

Contents

1. What robot builders buy, and why
2. The prices on record
3. Why bulk footage collapsed, and what still sells
4. Who pays
5. An existing archive, or recording to order
6. Consent, faces and the workplace
7. What makes a dataset worth more
8. How a sale through AIData.Broker works
9. Questions people ask
10. What this means for you

1. What robot builders buy, and why

A language model learns from text the web already held. A robot has no such web: it learns to grasp, carry, fold and assemble from video of people doing those things and from recordings of robots being steered through them. Builders buy three kinds of data.

• First-person (egocentric) video. A camera on the head or chest records a person's hands doing a task, ideally with the motion sensors and depth that tell a model where the hands are. Public research sets exist, such as Ego4D's 3,670 hours, but they are small against what builders now collect.
• Teleoperation demonstrations. A person steers a real robot through a task while every joint position, camera frame and command is logged. It is the closest data to what the robot itself will do, and the slowest to make: one Chinese account says that a year earlier a team working eight hours with a robot often produced one valid hour of data.
• Footage of real workplaces. Kitchens, warehouses, factories and vehicles hold the tasks robots are being built for, and they are the places a collection vendor cannot film without the business's agreement.

The scale of collection shows the demand. Figure's Index app reports $35 million paid to contributors, 28.5 million video uploads and contributors in 111 countries. GoPro reported more than 500,000 hours of video submitted to its AI training programme, and one recorder programme collected 38,000 hours from 3,900 contributors by mid-summer. What labs buy shows where this kind sits among the others.

2. The prices on record

The prices for physical-world data spread across the whole value chain, from what a recorder is paid to what a finished dataset asks. Read them by side: what the maker is paid, what a buyer pays a vendor to record to order, what a finished dataset asks, and what sold.

Teleoperation is priced apart, per operator hour: one vendor's guide puts it at $50 to $200, and in China real-robot teleoperation data is quoted at about 500 to 1,000 yuan per hour of data. One robot data vendor also sells it by the episode, from $5 to $15 for a humanoid or dexterous episode recorded to order.

Per hour of video, buyers pay 2.7x what the people who record it are paid.

3. Why bulk footage collapsed, and what still sells

First-person footage was scarce in 2025 and is not now. Anyone with a head-mounted camera can record household chores, collection programmes have spread to many countries, and in India collection rates fell about 30% in the first half of 2026 as supply grew. Buyers pay for accepted hours, and one collection vendor discards part of its raw capture in its own quality checks. In India the industry is weighing a common rate of $7 to $10 an hour for collectors, against far less paid today.

What still sells at a price falls into four groups.

• Teleoperation of real robots, because it needs the robot, the operator and the floor, and yields slowly.
• Footage captured to a buyer's specification: the tasks, camera position, sensors and labels the buyer asked for, which is why our valuation prices spec footage at $15 to $40 an hour against $1 to $10 an hour for unsorted footage.
• Rare settings: industrial lines, commercial kitchens, warehouses, hospitals, vehicles, where a vendor cannot send a recorder without the operator's agreement.
• Multiple sensors and views: wrist and third-person cameras, depth and force logging added to the head camera, which one vendor prices at $200 to $450 a captured hour against $80 to $200 for head camera alone.

Simulated recordings are a different product. Our valuation puts no price on footage a model generated: buyers pay for footage people made, and the one generated set on our board is only listed, not sold.

4. Who pays

Five buyers on our board take physical-world data, and they work in different ways.

Beyond those, collection platforms hire people to record by the hour: OpenTrain AI posts household and vehicle recording jobs, and micro1 has offered $15 to $20 an hour of approved video for car-mounted footage. Those are jobs for individuals, not purchases of existing data. The buyers page lists every buyer with the rules it publishes.

5. An existing archive, or recording to order

Most offers a business hears are offers to record: a vendor brings cameras, pays or covers the workers, and owns or licenses what is recorded. That is a service the business provides, priced by the hour of footage or the hour of work.

An existing archive is different, and rarer. A robotics company's teleoperation logs from its own development, a security or quality-control archive of a production line, a training library of procedures filmed for staff: these already exist, cannot be recorded again without the original setting, and can be sold in one sale to the buyers they fit. They are what AIData.Broker represents. Before valuing one, check three things:

• Who appears in it, and what they agreed to. Section 6.
• Whether the footage is the business's own. Footage a vendor recorded under its contract may belong to the vendor or be licensed exclusively to it; read that contract first.
• What the data carries beyond the picture. Joint logs, sensor streams and task labels are worth more than the video alone (section 7).

Value the footage: the valuation asks how many hours and what kind, and returns a range from the public prices on the board.

6. Consent, faces and the workplace

Physical-world footage is footage of people. Faces and voices can be biometric identifiers, and Illinois's biometric privacy law says that no private entity holding one " may sell, lease, trade, or otherwise profit from" it. That ban has no consent exception, and whether raw video holds such identifiers is contested, so a cautious buyer wants written releases naming AI training from everyone identifiable and a promise that no face or voice templates will be extracted, or footage where nobody can be identified, and so do we. Footage with sound can also fall under all-party consent laws for recording, such as California's. We do not take recordings of people who did not agree to that use to market, and our de-identification standard lists the rule with the law behind it.

In a workplace that has three consequences.

• Employees' agreement is part of the archive. A recording made for safety, training or quality control was made for that purpose; licensing it for AI training is a new use the people in it did not agree to. Without their written agreement, the footage is valued only where nobody can be identified.
• Customers and visitors count too. A kitchen or a shop floor films more than staff.
• Footage that avoids people is easier. Hands, tools and objects in frame, faces out of it, is what most collection specifications ask for anyway.

7. What makes a dataset worth more

Buyers pay for the things that save them work or that they cannot make themselves.

• Labels for what is happening: the task, its steps, success or failure. One narrated egocentric set, whose narration says what each clip shows, listed $488,200 for a 50,000-hour batch on 11 October.
• Sensors in sync: head and wrist cameras, depth, motion sensors and, for robots, joint and gripper logs, all time-aligned.
• Coverage: many people, settings and objects rather than one person repeating one chore.
• Clean rights: written releases, the business's own footage, no vendor's exclusive licence over it.
• Rarity of setting: an industrial line, a commercial kitchen or a fleet a buyer cannot film.

8. How a sale through AIData.Broker works

AIData.Broker represents holders selling or licensing their data to AI companies, a business's footage and a robotics team's logs among them. We work for the holder only: we value the data against public prices, settle its rights with the holder, and put it before every buyer that fits at once, so the holder sees competing offers rather than the first one. The holder itself licenses or sells to the buyer, and we do not take ownership of the data.

1. A valuation first. The valuation asks a few questions (how many hours, what kind, who appears in it) and returns a range from the public prices on the board, with no upload.
2. The rights file and the datasheet. We go through the rights questions with the holder (releases, ownership, vendor contracts), record the answers, and write the datasheet that describes the data; the holder approves it before any buyer sees it.
3. The round. A blind profile goes to every buyer that fits, without the holder's name, less any buyer the holder leaves out. Buyers who want to go further sign an undertaking and see the datasheet and, where the holder agrees, a sample.
4. Offers and the licence. Buyers make offers, which we set side by side; the offer check reads any written offer against market terms. The holder chooses, and signs the licence with the buyer directly.

The steps and documents are on how it works. Our fee is a success fee on what the holder is actually paid, band by band: 12% of the first $250,000, 10% up to $1 million, 8% up to $3 million and 6% above $3 million, and nothing if nothing sells; on a $300,000 sale it comes to $35,000. The mandate is not exclusive: the fee is due only on a sale to a buyer we introduced, during the mandate and, if that buyer is on the list we send in writing when it ends, within 12 months after it. We run sales only for holders based in the United States, the United Kingdom and Canada.

Which data fits. The smallest sale we take on is $23,000. At the prices above, that means a substantial archive: thousands of hours of spec or workplace footage with releases, or teleoperation logs of real robots. A small library of household-chore footage is better sold through a collection platform's own programme.

Questions people ask

Who pays for egocentric video data? Robotics companies and the vendors that collect for them. Figure reports $35 million paid to contributors through its Index app; collection platforms pay recorders a median of $14 per hour of video; vendors recording to order charge $38 per hour of video. Buyers of existing libraries, such as Sieve, pay after delivery is checked.

How much is an hour of first-person video worth? It depends on what it is. First-person footage of manual work is priced at around $1 to $10 an hour by one Indian collection startup, by task complexity, footage captured to a specification at $15 to $40 an hour in our valuation, and multi-sensor capture to order at $200 to $450 a captured hour.

How much does teleoperation data cost? One vendor's guide puts teleoperation at $50 to $200 per operator hour, the quote in China is about 500 to 1,000 yuan per hour of data, and episodes recorded to order run from $5 to $15 for humanoid or dexterous tasks. Our valuation prices existing teleoperation demonstrations at $28 to $60 an hour.

Can a company sell its CCTV or training videos to robotics companies? Through us, only with the written agreement of the people who appear in them, or where nobody can be identified. The footage was recorded for safety or training; selling it for AI training is a new use, and face geometry or voiceprints taken from it can be biometric data under laws such as Illinois's.

Why did prices for egocentric video fall? Supply grew faster than demand for unsorted footage: collection spread to many countries, rates in India fell about 30% in the first half of 2026, and buyers pay only for the hours they accept. Teleoperation and multi-sensor capture still cost several times bulk footage.

What this means for you

Robot builders are among the most active buyers of new data, but their money goes mostly to recording, not to archives. A holder earns most where its data cannot be recorded again: real robots, real workplaces, the sensors in sync and the people's agreement in writing.

1. Know which kind you hold. Bulk first-person footage, footage to a specification, workplace footage or teleoperation logs: bulk footage is priced at about a tenth of footage to a specification, and teleoperation above both. The rates page and the physical-world data page hold every public price.
2. Settle consent first. Written releases naming AI training from everyone identifiable, with no face or voice templates extracted, or footage where nobody is identifiable.
3. Check who owns the footage. Your own recording, or a vendor's under its contract.
4. Keep what makes it worth more. Sensor logs, labels, the task each clip shows.
5. Value it, then put it before every buyer at once. Value the footage, and read any offer against the market with the offer check.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. The holder pays us a fee fixed in the mandate and paid only if a sale closes, and we run sales only for holders based in the United States, the United Kingdom and Canada.