Why is a robotics company paying $15 an hour to film someone folding laundry?

TL;DR: Text was relatively free to scrape. Movement is not, so robotics companies are buying it, from people doing the job. Figure has paid $15m to people filming their own chores and has committed over $1bn to data and compute over the next twelve months.

Recorded human labour has quietly become a tradeable asset. If you already film your own people for quality, safety or training, that asset is sitting in your building.

The robotics company Figure will send someone to your home to do your chores. They arrive wearing a camera. The app that arranges it also lets you sign up the other way round, to be the person who records, and it will send someone to your business too.

The chores are not the product. The footage is.

That lands differently if your own business already films people working, for safety, for training, or because of a dispute nobody wants repeating. Same category of thing.

Language models got their training data for nothing. The internet was already written down, already public, already sitting there. Robotics has no equivalent. Nobody uploaded a video of themselves unloading a dishwasher with the camera on their forehead and their hands in frame, because why would they.

So it is being bought. Figure's app, called Index, came out of stealth on 25 August. It has paid $15m to contributors so far, runs in 108 countries with about 44,000 people recording each week, and processes 30 minutes of uploaded video every second. Figure has committed over $1bn to data and compute over the next twelve months.

Interesting fact: they tried the normal route first but failed. In Figure's own words, they tried buying the data, and no vendor could hit the volume or the variety. So they went to individuals instead.

Not for the reason most people assume. Seeing is basically sorted. Walking is basically sorted, which is why every humanoid demo you have seen involves a backflip. The gap is physical touch.

Carrying a full mug across a dark room takes a few hundred small corrections in your fingers, corrections you could not describe if somebody asked.

A human hand has around 17,000 touch receptors and can feel a difference in texture measured in nanometres. A Unitree robot hand has 94 sensors. The Shadow hand, billed as the most advanced five-fingered robot hand in the world, costs about $100,000 and lifts 4kg. Simulation does not rescue this either: it transfers well for walking, where gravity is the dominant force, and falls apart for anything involving friction, give and slip.

MIT Technology Review followed the workers at Micro1, which has thousands of contractors in more than 50 countries strapping phones to their heads and filming themselves folding laundry and washing up. The rate is $15/hour, which is good money locally. However, none of the workers interviewed knew who ends up with the footage, and the company will not say, citing client confidentiality.

If we skip all the confidentiality risks, every second factory now produces two things of value, a physical product for the employer and a reusable training dataset for a completely separate market. How many workers know about both parts?

The industry is not fully aligned here. For example, Ken Goldberg at Berkeley has two papers in Science Robotics arguing that you cannot collect your way out. His figure for it: all the text used to train today's language models would take a person about 100,000 years to read, and robots need more than that, not less. His alternative is to ship something that works, then let it generate data in production. His own sorting robots ran four years and produced 22 years' worth of real picking data.

There is a sharper number on his side. A Stanford and Berkeley team took a T-shirt folding robot from 8% to 83% success on the same 200 hours of demonstrations, by training a second model to score which bits of that footage were any good and weighting the training accordingly. Not more data. Better use of the same data.

Either way the footage is being collected, so the rights question stands whoever wins that argument.

The ICO's guidance on monitoring workers explicitly covers wearable cameras and body-worn devices, and it applies to gig arrangements too, regardless of the nature of the contract. You must document why you are monitoring and what you intend to do with what you collect. If people work from home, your assessment has to account for household members caught in shot. And there is one line in it that reads differently this year: you should not assume that packages you purchase are compliant with data protection law.

The TUC has gone further. Its model AI Bill, drafted with Cloisters and Cambridge's Minderoo Centre, would give unions a legal right to the data an employer holds on their members. A separate TUC paper queries whether genuine consent is obtainable inside an employment contract at all. That last point is not only a union position. Treating consent as unreliable where one party depends on the other for their job is ordinary UK GDPR doctrine, which is why employers are usually advised to rely on a different lawful basis.

Somewhere in your organisation there is probably footage of people doing skilled physical work. Filmed for training, or safety, or a quality dispute. Nobody thought of it as an asset because until recently it was not one.

Worth writing down what that footage may and may not be used for, and by whom, while the answer is still yours to decide.

P.S. The government's consultation on who may act on people's data rights closes on 7 September.