India's Human Robot Trainers Fuel Global AI Race

India is seen as a laggard in the global AI race, but having the world's second-largest workforce and relatively low labor costs has led to a boom in companies collecting data to train robots. I spoke to the workers and companies in India that are fueling this trend of humans indirectly training robots.

When it comes to making AI-enabled robots, China and the U.S. are miles ahead of India. New Delhi is seen as a laggard in the global AI race and, by extension, in the physical AI space.

But having the world's second-largest workforce and relatively low labor costs has given India a unique advantage — it can provide the humans to train robots.

Tanisha Reddy, a teacher at a private school in southern India, moonlights as a robot trainer. She starts the day by recording first-person videos of herself performing routine tasks like cooking, cleaning dishes, and packing lunch, she told CNBC in a phonecall. In the evening, she repeats the process, generating 3-4 hours worth of video every day. She gets paid less than $4 for every hour of recording.

"I am very happy," said Reddy as she explained that the work is simple and takes no extra effort, and does not involve her taking time away from raising her two children.

It has been more than four months since she started working for Qanat Consulting Services, a firm in Andhra Pradesh.

In less than a year, several firms have cropped up in India that recruit people to record first-person or egocentric videos. Many of them record data, check quality parameters, and perform data annotation on behalf of clients in the U.S. and China. These could be either robotics companies or, at times, other intermediaries.

"We get contracts mostly from companies in the U.S. and China," Thaslim Pattan, founder of Qanat Consulting Services, told CNBC, adding that she recently won a contract to source videos from workers in a garment manufacturing company.

Robots are developed in a lab, but they need to be trained for a real-world environment, she said. With the market for robots expected to grow rapidly over the years, the demand for data collection is soaring.

Global brokerages are bullish on the growing demand for robots, with Barclays expecting the humanoid robot market to grow to $200 billion in less than a decade. At the same time, Morgan Stanley forecast it would surpass $5 trillion by 2050, with the number of robots reaching 1 billion.

But despite the rising volume of work, Pattan said contract prices for data collectors are falling as the number of competitors is growing. Within months, prices have already halved, she added.

Data collection is being commoditized, and Indian companies in the space will need to move higher up the value chain to remain relevant, experts said.

Retaining the advantage

Neocambrian AI, a north India-based startup, last month launched a robotics data factory in Noida where it is collecting robotics data in simulated environments, Abhinav Kukreja, the firm's founder, told CNBC.

He has also established a network of more than 100 factories where workers record themselves performing tasks.

The firm is directing efforts towards creating datasets that are "useful for solving dexterity," and teaching robots "how to manipulate objects," he said, adding it would require 100 million hours of video to reach a human-like level of dexterity.

For example, a robot needs to comprehend the different levels of pressure that are required while holding an egg versus holding a bottle of water. Every additional dataset helps the robot improve its ability to manipulate objects like a human hand.

His firm's core business is to process data, and unlike many of its peers, it maintains ownership of it. "We don't go to customers and ask them what they want," explained the founder, adding that instead it provides customers with datasets that Neocambrian AI has pre-emptively built.

"Across the AI stack, this is the only layer where India can not only participate but win," said Kukreja. Human data has been valued ever since AI has existed, and India could become a "human labor marketplace of the world," similar to its experience in the information technology sector.

Experts tell me that much like smartphones, robots have a hardware aspect and an operating system. While India has a long road ahead when it comes to manufacturing robots, they said the country could play a role in developing the operating systems.

Humyn Labs, another Indian start-up in the space, is focusing its efforts on data conversion and building a diversified data set from different environments. It claims to offer "(verified) human intelligence at scale" and collects 50% of its data from Latin America, 35% from India, and 15% from elsewhere in Asia.

Its focus is on converting and owning the datasets, Manish Agarwal, co-founder of Humyn Labs, told CNBC.

It is the early days of training robots using videos, and eventually, the market for data collection work will become saturated. India needs to "evolve from collector to converter" to keep its advantage, he added.