Workers are filming their hands to teach robots. The robots have a long way to go.

Shoba Kirubakaran has worked at a textile factory in Karur, in southern India, for ten years, six days a week. At the moment she packs aprons: yellow ones, with chickens on, each wrapped in plastic for a homeware brand. She earns about $90 a month.

While she does it, there’s a GoPro strapped to her forehead.

The camera is recording her hands: how they pick up an apron, fold it, slide it into its wrapper and set it on the pile. The footage is for training robots, so that one day a machine can do what she does. She told NPR, which reported her story today, that nobody at the factory had said so. She found out from the reporter.

She’s far from the only one, and the reasons robot makers want this footage say a lot about where robots are, and aren’t, right now.

Hundreds of hands, on camera

In April a video of rows of garment workers in head cameras went round Indian social media. Scroll traced it to a factory in Gurugram, near Delhi, where workers said they’d been asked to wear the cameras through their shifts without anyone asking their consent. India’s oldest trade union said it added to the exploitation workers already face. The OECD, which keeps a register of AI incidents, logged it as a hazard: no harm yet, but plenty of worry.

The cameras come from data companies, which sell the footage on to the people building robots. In May, AFP found eight workers at a textile factory in Karur wearing head cameras and smart glasses supplied by one of them, Objectways. AFP didn’t name the factory. Data companies also pay people to film themselves at home. In Chennai, AFP met a woman recording herself slicing mangoes, paid 250 rupees, about $2.60, for each hour of footage. Folding clothes, making coffee and making sandwiches are among the jobs clients ask for, Objectways’ chief executive told AFP.

The Los Angeles Times met one of Objectways’ staff, Naveen Kumar, 28, an engineering graduate who straps a GoPro to his forehead and films himself folding hand towels, following a strict list of hand movements. His firm had sent 200 towel-folding videos to a client in the US.

Not all of it is aimed at somebody’s job. NPR reports that some of the clients in India want robots for work that’s hard to find people for: tapping rubber, dusting solar panels out in the countryside, pulling the stems off chillies.

Why a robot needs to watch

The AI that writes and talks learned from the internet: an enormous pile of text, and of pictures with words next to them. Nothing like that exists for hands. There’s no internet of people folding towels, filmed from where their eyes are, showing exactly where their fingers go.

That’s what this footage is. The industry calls it egocentric data, which just means first-person video, shot from the worker’s own eye line. A camera on a tripod sees somebody fold a towel. A camera on their forehead sees what they see while they do it: which corner they reach for, where the thumb goes, what happens when the cloth slips.

This site has written about towels before. In our post on Moravec’s paradox, the robot couldn’t fold one, which is the paradox in a nutshell: the things computers find hard are the things we do without thinking. Some robots can fold a towel in a demo now. Doing it all day, with any towel, in any room, is what the footage is for.

How much footage is that?

Ken Goldberg, a roboticist at the University of California, Berkeley, has a way of putting the size of the problem. Take the text and images that today’s big AI models learned from, turn them into the time a person would need to read and look through them all, and you get something like 100,000 years. He calls the difference between that and what robots have to learn from the “100,000-year data gap”.

So how big is a pile of footage, next to that? Pick a stop and the chart zooms to it.

How much footage is that?

Hours of first-person video, drawn as squares. A square's area is its hours, so a bigger amount swallows the smaller ones.

1. One hour: what a data company pays about 250 rupees (US$2.60) for, filmed at home.
2. One shift: a working day, on camera.
3. Ego4D, 2022: the first library of first-person video on this scale, from 931 people in nine countries.
4. Egocentric-10K, 2025: footage from 2,138 factory workers, released under an open licence.
5. One worker's decade: ten years of six-day weeks, if the days are eight hours long. Our estimate.
6. Egocentric-100K, 2025: from 14,228 factory workers. About eleven years of footage, played end to end.
7. 100,000 years: roughly how long the data behind today's big AI models would take a person to read and look through, by roboticist Ken Goldberg's estimate.

The big collections of first-person footage are big by any human measure. Ego4D, put together by thirteen universities and Meta, has 3,670 hours, filmed by 931 people in nine countries. A start-up called Build AI released Egocentric-10K last year, 10,000 hours of footage from 2,138 factory workers, under an open licence, and then Egocentric-100K: about 100,000 hours, from 14,228 workers.

Now put one person beside them. If Shoba’s days are eight hours long, her ten years at the factory come to about 25,000 hours. That’s two and a half times Egocentric-10K, from one woman, at one table.

And on Goldberg’s scale, the biggest of the datasets is a dot. You’d need about 8,700 of them to fill the 100,000 years.

It isn’t an exact comparison, because Goldberg’s figure turns words and pictures into time and that’s not the same thing as hours of video. He also doesn’t think watching people is the only way to close the gap. Robots can practise in simulations, learn while a person steers them, and learn on the job, and his own bet is on getting robots earning their keep at narrow jobs first, so that they gather data as they work. But the scale explains the rush. Hundreds of people filming their hands is how you start on a gap that size.

Which hands are on camera

Every job on this site has two numbers that matter here, both from O*NET, the US government’s big database of what jobs involve. One says how hands-on a job is. The other, which we call fiddly, combines how much finger and hand dexterity the work needs with how often it’s done in cramped or awkward spaces. Carl Frey and Michael Osborne picked those three, in their well-known study of which jobs computers could take over, as their measure of the handwork machines find hardest. There’s more on why in Hands-on is not one thing.

Here’s every job we can measure, with the ones now on camera lit up. You can light up other groups, or find your own job.

The cameras are pointed at the tidier end of hands-on work

419 jobs by how hands-on the work is (across) and how fiddly the handwork is (up), both from O*NET, out of 100. Tap or hover a dot to see the job.

The map needs JavaScript to draw. The table below has the jobs it's about.

The jobs on camera are bunched together, and they’re all below the same line. This site counts a job as hands-on if it scores 50 or more on the first number, and 195 jobs do. Half of those come in above 39 for fiddliness, and half below. Every job on camera is below. Hand packers, Shoba’s job, rank 166th of the 195.

Light up some of the fiddliest trades instead and they’re at the other end: electricians, mechanics and plumbers, working with small parts in tight spaces. None of those jobs turn up in any of the footage we’ve read about.

Job	What’s being filmed	Hands-on	Fiddly	Our score
Hand packers	Wrapping aprons, in Karur	59	33	80
Sewing machine operators	The garment factory in April’s video	51	36	85
Hand launderers and pressers	Folding towels	61	35	86
Domestic cleaners and helpers	Housework, filmed at home	44	22	86
Kitchen helpers	Slicing, sandwiches, coffee	42	31	87
Manufacturing labourers	Sorting and packing, in the factory datasets	65	38	88
Assemblers	Assembly, in the factory datasets	61	36	68
Building electricians	Nothing we’ve seen	76	56	81
Motor vehicle mechanics	Nothing we’ve seen	71	53	82
Plumbers	Nothing we’ve seen	66	50	85

That fits what robots are like. They do best in tidy, predictable places, doing the same thing over and over. Packing the same apron into the same wrapper at the same table, thousands of times, is about the best classroom a robot could ask for. A crawlspace under a kitchen floor is about the worst. The cameras are pointed where robots have the best chance, which makes sense, and which is also exactly where these jobs are.

One oddity worth knowing: the US data rates domestic cleaners and kitchen helpers as a little under half hands-on, which may surprise anyone who’s done either job. It’s measuring how much the physical side matters next to everything else the job involves, and it puts both just left of the line on the map.

Now look at the last column. These jobs score between 68 and 88 on this site, which means most of the work still needs a person. That’s true, and it’s also beside the point here. Our scores measure what AI software could do with each task, using the International Labour Organization’s ratings, and software can’t wrap an apron. Robots aren’t in the number. Every hands-on job page here says so, and says that for work like this, robots are the bigger question. We learnt that one the hard way. This is what the bigger question looks like when somebody starts working on the answer.

Where the robots are now

None of which means a robot is about to take Shoba’s place at the table.

The research firm IDC counts about 25,000 humanoid robots shipped around the world in the first half of this year, more than five times as many as a year before, and more than three quarters of them in China. But most of them went somewhere to be studied, taught with or shown off.

Most of this year's humanoid robots went somewhere to be studied or shown off

Humanoid robots shipped worldwide from January to June 2026, about 25,000 in all. Tap a robot.

• 69 in 100: research and teaching, shows and displays, and government data centres.
• 31 in 100: everywhere else, factories and warehouses included.

Of the robots shipped in those six months, fewer than 8,000 went anywhere else. Where humanoids are working, it’s in small numbers. This week South Korea cleared 20 of them to work at chip, car, shipbuilding and logistics sites, and about 20 are loading and unloading at a car-parts factory in Hangzhou. They do supervised work, and they can’t yet switch from one job to another or be left to get on with it.

The Los Angeles Times put it neatly: the robots can be impressive when somebody else is controlling them, but they’re still a long way from working on their own. The footage is a bet that this changes.

Whose footage is it?

The part of this story that’s already settled isn’t about robots. It’s about the footage.

What’s being recorded is skill. Ten years of packing aprons is what makes Shoba’s hands quick and sure, and that’s what a robot maker wants on film. In her case, and in the Gurugram factory’s, it was collected without the workers being told what it was for.

It doesn’t have to be done that way. When the universities behind Ego4D collected their 3,670 hours, the people wearing the cameras had agreed to it, and faces and other identifying details were blurred where needed. Nobody needs to invent the ethics of filming people’s hands. They already exist, and some of the people doing the filming are using them.

What it means for you

If your job is hands-on and repetitive, our score won’t show you robots, and it says so. Look at the fiddly number on your job’s page: the lower it is, the more your work looks like the jobs being filmed. That’s not a countdown. Robots are a long way behind the footage, and further behind the demos. But it’s a good reason to get good at the parts of your job that change from day to day: sorting out what’s gone wrong, handling the odd one out, showing the new starter how it’s done.

If someone hands you a camera, ask what the footage is for, who’ll see it, whether you can say no, and whether you’ll be paid for it. They’re fair questions anywhere, and a decent employer will have the answers ready.

If your work is fiddly, in cramped spaces, with small parts and no two days the same, you’re at the far end of the map from the cameras. That’s the hardest place for a robot to learn, and nobody is filming it yet.

Robots will probably learn to fold towels properly one day, and when they do, it’ll be from footage like this: thousands of people who are very good at folding towels, showing them how. That seems like something those people should be asked about, and paid for.