印度成AI机器人数据收集中心，美企高薪招募劳工

Bloomberg Audio Studios, podcasts, radio, news. On top of a mountain of discarded plastic, a 43-year-old woman named Senita Ratau is hard at work. She's in a bright yellow sari with green glass bangles sifting through piles of litter, plastic waste, which she'll sort and clean for recycling. Senita is based in New Delhi, but scenes like this one take place every day all over India on construction sites, in homes and recycling plants where manual labor is done.

Nothing out of the ordinary except for one thing.

Suddenly, you saw this high-tech contraption on her forehead, an iPhone.

An iPhone strapped to her head recording every move she makes. Stha Ry covers tech in India for Bloomberg. She met Senita at a recycling plant earlier this year. Senita has spent years sorting and cleaning discarded plastic for about $200 a month. Usually the factories where she works are lively places.

There's a lot of banter. There's a lot of conversation. There's a lot of silliness that goes on. in this. It was deadly quiet because all these women were not allowed to talk to each other

8 hours a day in total silence. But that's not all.

She was hunched forward constantly because she was not loved to capture any other part of her body except her wrist. It's exactly what uh a humanoid robot camera would see if it looked down. That was the motion she was asked to capture. Senita doesn't know what the footage is for. All she knows is that this phone recording earns her extra money. Money she uses to send her four kids to school. But the recordings she and her fellow workers make are critical to the future of AI and robotics.

They'll be fed into AI systems to teach humanoid robots how to move and navigate the physical world like us, one frame at a time. The goal really is to capture the micro gestures, the really tiny hand movements that, you know, robots need to learn to become better at everyday tasks. The world has started to embrace chat bots and agentic AI systems like Claude or Chat GBT which rely on cloud computing or local hardware to run.

You give an AI agent a voice, text prompt or task, it gives you an answer. And for years, the robotics industry has tried to create humanoid devices. AI that moves in the physical world the same way that we do. When it's gone well, it's been dazzling.

Okay. Oh,

we have a new world record.

At the World Humanoid Robot Games in Beijing last month, essentially an Olympics for machines, a Chinese robot shattered the 100 meter sprint record. Its time 8.64 seconds. Nearly a full second faster than Usain Bolt, the fastest man on Earth. But step outside that sterile controlled track, and it's a completely different story. When robots come into the real world, there's invariably a blooper reel awaiting the texture of real life, the unevenness of perhaps the terrain, the the spatial intelligence that's required by for these robots.

All of these things, you know, make it really difficult for robots to be trained in the lab. They have to be trained in the real world with real world data, which is where, you know, all of these data collectors come in. Cityroup projects the market for human-like robots could reach $7 trillion by 2050. And to learn how to perform tasks like a human, robots need huge amounts of so-called egocentric data. Footage captured directly through the eyes of a person.

China has been the biggest generator of that data, but it's keeping it within its borders. So the demand has shifted to the world's most populous country.

So India really has become the epicenter of data collection in the past few months. Tens of thousands of workers like Sonita uh shoe factory workers, denim factory workers, construction workers, all kinds of people have been recruited to record these firsterson camera footage or what is called egocentric footage. It was a sort of unsettling moment for me when I talked to these workers because most of them were recording themselves into their own job extinction.

Welcome to the big take Asia from Bloomberg News. I'm Rebecca Chung Wilkins in for one. Every week we take you inside some of the world's biggest and most powerful economies and the markets tycoons and businesses that drive this ever shifting region. Today on the show, why the road to the next AI breakthrough runs through India and what this race for human data means for the US China tech rivalry. Humanoid robots and online chat bots both use AI.

But Bloomberg Sera Ry says teaching a large language model to write an essay is completely different from teaching machine to cook a meal.

Real world humanoid robotics is a completely different ballgame. This requires real world data, actual human motion and human movements that have to be learned by the models to teach the robots. There is no availability of any of this data online, you know, on the internet.

You might think, wait, the internet has plenty of data. All those videos on YouTube or Tik Tok, all those humans doing random things. But

it's really raw data and it's not firsterson egocentric data. It is not the data that a humanoid robot's camera would see when it's looking down or looking forward. That kind of data has to be collected in a highly standardized way. Faces can't appear in the footage and conversations aren't allowed. Equally important, the data has to come from as many different humans as possible.

Each person has their own way of doing things. Like how you perhaps open a can be completely different from how I open a can. So each person's data is valuable and multiple labs are looking for what is called proprietary data.

Sura says India is an ideal place to gather this kind of data. The country has an enormous population. Labor is cheap and workers still do a lot of manufacturing by hand. the diversity whether it's the shoe factories, the denim uh factories, whether it's the you know warehouses of different kinds, whether it's the households, fishing villages or the farms or construction sites or the mines. It's incredible the diversity of tasks that can be recorded in India.

No other country could offer this kind of a diversity of manual labor. Not even China I would say because a lot of these tasks particularly in factories and warehouses are already automated in some way or the other in China.

And what about the cost? Can you break down what kind of cost you you have if you're buying this data in India versus somewhere else say Singapore or even the US?

I haven't looked at uh the costs in Singapore but I can definitely talk about the costs uh in India versus the US. For instance, the raw first person footage of the egocentric data that India provides could come for about $7 an hour. Uh but if it's cleaned and annotated, it could go for as much as $30 an hour. That's a huge value addition right there. And the same data, you know, could go from two to three to five times more expensive in the US when it's collected in the US.

Tesla, for instance, pays $48 an hour in Palo Alto for um motion capture data collector inhouse.

The factory floor is just the starting point for these egocentric videos. Once the footage is captured, it's sent to local firms that clean it up, check the quality, and label the data so it can be used to train AI models. Demand for this work has surged, helping fuel a boom in India's AI services sector with a growing number of companies built around processing and supplying data to tech firms. Who are these startups working with?

What's the sort of end point of all of this proprietary data?

Now, none of these companies were able to tell me who exactly was the end customer. Uh not that they didn't know, but the NDAs were so strict and so binding that none of them were allowed to reveal. But enough hints were dropped that these were all like major American and global companies including Google and Tesla and uh Nvidia and Meta. You know, Nvidia for instance is now getting into uh physical AI and embodied