In Karur, in southern India, Shoba Kirubakaran, 40, packs yellow aprons decorated with little hens. On her head, a GoPro camera records her every movement: the way she picks up the fabric, folds it, slips it into the plastic.
It is the third time she has been asked to wear it. What she has not been told, she tells American public radio NPR, is what these images are for: teaching robots to do her job.
A new market: our gestures
To learn to write, an AI can read billions of pages available on the internet. But to learn to fold an apron, tidy a kitchen or screw in a part, a robot needs something else: to see these gestures, up close, exactly as the person performing them sees them.
Specialists call this "egocentric" data: videos filmed in the first person, from a camera mounted on the head, chest or wrist. And India is becoming one of the world's major suppliers of it.
These videos are shot in textile factories, but also in private homes, hotels, restaurants, and even in fake kitchens and fake bathrooms set up in offices, where employees fold napkins and do the washing-up in front of the lens. The images are then cleaned, cut, annotated, and sold to companies that design humanoid robots, in the United States as well as China.
How much it pays
According to an AFP investigation, some workers earn around 250 rupees, roughly 2.60 dollars, for an hour of video. An American start-up, for its part, offers a base rate of one dollar an hour. And several investigations, notably by the Guardian, have shown that in some factories consent was given by management, not by the workers being filmed, many of whom did not know what the images would be used for. Some were not paid at all for this part of their work.
In April, a video showing rows of textile workers fitted with head-mounted cameras shocked Indian public opinion and circulated widely on social media.
In 1988, robotics researcher Hans Moravec made an observation that became famous: what is difficult for humans, such as playing chess or calculating, is relatively easy for a computer. But what seems child's play to us, grasping an object, walking, folding a garment, is extraordinarily difficult for a machine. We spent millions of years of evolution mastering our hands. That is why an AI beats the world chess champion, yet no robot can still properly put away your laundry.
Will robots really replace them?
Not right away, according to several specialists interviewed by NPR. The needs are gigantic: companies in the sector estimate that hundreds of millions, perhaps billions, of hours of video will be needed to train genuinely capable robots. And one business owner originally from Karur even doubts that a robot will one day cost less than an Indian worker.
We explained this in relation to the household robots that still do not exist: the human hand remains an immense challenge. It is also why companies like World Labs are trying to teach AI to understand space.
The workers, for their part, have their own idea. Kannagi Periyasaamy, 42, has drawn her conclusion: the work they do today will be done tomorrow by robots.
What we take away
For centuries, a skill was passed from one person to another through apprenticeship. Today, it is filmed, cut up and sold, often without the person who possesses it knowing.
Which raises a fundamental question, to which we devote our reflection this evening: who owns the know-how of our hands?