The Indian workers filming their work to train AI

Mohammed Ashim is paid to film his hands at work

Mohammed Ashim has been cutting hair and trimming beards for four years in Bengaluru.

But recently the 22-year old has been strapping a camera to his head to record his work.

The camera only records his hands and fingers - his client's faces are blurred out.

"The routine is simple, but you have to be disciplined about how you move.

The operations team taught me that my hands always have to stay cleanly inside the frame, the lighting has to be sharp, and I can't tilt my head too abruptly while working."

He wears the camera for five or six hours a day.

"It's good money, but honestly, the feeling of doing something this futuristic is what keeps me excited.

I never imagined my razor techniques would be useful to a global tech company," says Ashim.

He is one of hundreds of thousands of workers across the world paid to record their activities.

Ashim's footage is collected by a local company, which might have hundreds of people filming their work.

That video is then sold on to technology companies and research organisations who use it to train artificial intelligence - in particular so-called physical AI, a type of artificial intelligence that can control robots in the real world.

Rashdan Technologies collects and processes such human generated data.

"Every robotics company, every frontier AI lab, has a specific kind of data they want: industrial tasks, agriculture, household chores, small business work like pottery, garment-making, embroidery, candle-making.

We go wherever the task is," says co-founder and CEO Shahil Baig.

His company pays $5 to $10 (£3.74; £7.47) for eight to 10 hours of recorded work.

"For a lot of these people, that's meaningful," Baig says.

Making sure workers give their consent and are comfortable with the process is important, he says.

"We record hands, not faces, and where multi-camera rigs do capture more of the body or environment, we blur faces out.

No names, no contact details, nothing that identifies them personally goes into what we hand off."

But his customers are getting more demanding.

"Now clients are asking for four to six cameras rigged around a person's head to cover the entire scene, not just the hands.

"Getting someone to agree to one small camera is one conversation.

Getting them to agree to wearing what's effectively a five-camera headset is a completely different, and harder, one.

We pay more for that tier of data collection because of what we're asking of people."

And the industry is becoming more crowded, Agarwal expects dozens more competitors to emerge in India this year.

Shahil Baig expects dozens more firms to start collecting video of workers

If you want robots to navigate the real world, they can't just be trained inside computer simulations, argues Manish Agarwal, the founder of Humyn Labs, which collects and processes human-generated data.

Most of that data comes from video, but sometimes collectors wear hardware on their hands and fingers to capture physical contact.

"Physical tasks require high-quality signals involving limb movements, fingers, wrists, elbows, shoulders and dynamic force application.

For example: applying the correct grip force so a plastic bottle doesn't crumble versus a steel glass," he says.

And until recently, that data was not available.

"There is no existing internet archive of embodied human activity - no dataset of a kitchen being cleaned, or a car being assembled by hands that know exactly how much force to apply.

That data has to be captured from scratch, in the real world, across diverse environments," says Agarwal, whose company has bases in Bangalore and San Francisco.

At the moment his firm has 96,000 data collectors across countries in Latin America and Southeast Asia.

Humyn Labs filters the thousands of hours of video supplied by those collectors, selecting the most useful content.

It's then labelled so the AI knows what it is looking at and packaged in a format that can be used by clients.

"You need to source across a variety of factories, a variety of tasks, in different geographies because if you want to build deployment for US factories or US houses, you'd want environments that are a proxy for that.

Residential houses in India may not be the right context for building a robotics use case for the US."

The factory workers teaching robots how to work like humans