companies · public sources · compiled 12 Oct 2026
Who pays for robot-learning data, and who collects it
A map of the physical-AI data market: the labs and robot makers that buy, the brokers in the middle, the field teams and gig apps that collect, and the hardware and checking tools around them. Prices are given in dollars per hour at each step. Vendor prices are their own claims, and each claim about a buyer is marked by how sure it is.
Findings across the market
$2–3 /hrRaw egocentric video, down from $6–10 a year agoBuild AI released about a million hours for free. Bulk head-camera footage now sells below what it costs to film ($5–20/hr).
44k weeklyContributors on Figure's own data appFigure built Index because vendors "couldn't hit the throughput, diversity, or quality bar": 16M videos, $15M paid out. Tesla and Sunday also collect in-house.
70+State-backed robot training centres in ChinaAnother 40+ are planned. Teleop data trades there at ¥500–1,000 per hour. That's a market price, not a subsidy.
0Labs that publicly confirm buying from a brokerEvery named buyer–vendor pairing comes from the vendor's own marketing or a leak. Most labs run two vendors plus an internal pool.
How the money moves buyer to worker, with the price at each step
Sources: Business Times (Humyn Labs), Flikforge, DreamVu, Mercor's robotics page, TechCrunch (micro1), Economic Times, Inc42, MIT Technology Review. Vendor figures are their own claims.
Price ladder US$ per hour, log scale
What buyers pay for dataWhat collectors pay people or sites
Ranges from vendor price lists (DataX Power, nxted, Claru, RCSV), press (Business Times, The Economist, TechCrunch, Economic Times, MIT Technology Review, Semafor) and job posts (Tesla). China teleop is a market price for data.
Who buys egocentric data and how sure we are
- Robot foundation-model labsPhysical Intelligence, Generalist, Skild, NVIDIA GEAR and Dyna. Scale names PI, Generalist, Cobot and Dyna as customersS. Generalist pretrains on 500k+ hours of wearable data and contracts partners per usable unit. Skild runs a Bengaluru vendor program.
- Humanoid makersFigure (mostly in-house), Tesla (in-house helmet rigs), and 1X (Mecka launch partnerS). Agility, Boston Dynamics, UBTECH and Unitree are named by Indian collectorsR, unconfirmed by the companies.
- World-model labsNVIDIA Cosmos and 1X's world model. Odyssey runs its own capture rigs. Lightwheel names Google DeepMindS.
- Industrial investors as buyersSamsung, Hyundai, LG and SKT both fund Config and buy from itR. Appen says China is its largest robotics-data market and the US its second.
- Not LLM labsNo LLM lab is named buying egocentric video for language models. The "frontier lab" buyers vendors cite are robotics or world-model teams. OpenAI's robotics lab collects its own teleop with about 100 collectors.
Where the value is moving what holds a price and what doesn't
- Commodity: footage a crowd can filmHead-camera and phone video with no special sensor, no robot and no action labels. Labs bring it in-house, crowds of tens of thousands supply it, and open releases undercut it.
- Holds value: the buyer's own deviceDemos on a buyer-specified gripper or glove, or teleop on the buyer's robot ($50–200/hr). The device belongs to the buyer, so a crowd can't self-serve it.
- Holds value: contact and forceTactile gloves and force channels a camera can't see. Funding is concentrated in China: PaXini has raised over RMB 4B, and Tashan, Xense and Daimon all raised in 2026. Tacta ($75M) and 6thSense (YC) collect it in factories.
- Holds value: proof it trainsEpisode verification, curation and real-robot evals. This layer has the fewest commercial players, mostly papers and small teams (Antioch, Robocurve, CosmicBrain).
- Crowded: India collectionHuman Archive, Build AI/Egolab, Humyn, Awign, Objectways, Neo Cambrian, FPV Labs and Instawork all collect there. Pay is ₹100–400/hr, and the press is focused on worker consent.
Directory companies in seven layers
Layer
Region