Why the robot data gold rush is on borrowed time

A NEW side hustle has been making its way around South-east Asia, one that doesn’t require learning a new skill.

Workers – whether they chop vegetables, pack boxes or assemble laptops for a living – are strapping cameras to their heads and recording themselves doing their jobs. These videos are then sold to data firms, who in turn sell them to robotics companies which use these clips to train robots to do the same task.

On paper, this seems like a good deal for the workers. Siddharth Kumar, co-founder and CEO of Jakarta-based outsourcing company MyRobin, said a worker who records five to six hours a day could earn up to 3.3 million rupiah a month (about US$187). That’s roughly 50 to 60 per cent of Jakarta’s minimum wage, which they can earn on top of the salary from their day jobs.

“You’re not doing anything extra,” Kumar noted. “It’s just good pocket money.”

The demand is coming from a simple gap in the robotics race: Robots need to learn how humans move through physical tasks, but much of that first-person data doesn’t exist online.

This has created a scramble for egocentric data, or video recorded from a person’s point of view. Robotics labs used this data to train vision-language-action models, which help robots understand what they are seeing and decide how to move.

The buyers include companies building robot “operating systems”, robotics firms involved in both hardware and software, and manufacturers preparing robots for specific factory or warehouse environments. That’s according to Manish Agarwal, whose Bengaluru-based company Humyn Labs works on collecting and preparing physical artificial intelligence training data.

But the boom may have a built-in expiration date. While egocentric video is cheap to make and missing from existing datasets, its usefulness in training is questionable. In addition, these videos only teach robots based on what people can see, and sight can teach machines only so much.

In physical work, for instance, it’s important to know how much pressure the worker uses to hold an object without crushing it.

SEE ALSO

Samsung creates new robotics division; plans research hubs in US, China and Japan

Albert Causo, co-founder and CEO of Singapore-based Hand Plus Robotics, said that most of the time, vision only helps robots identify where an object is, how to approach it, and whether the task is complete. That means in one hour of raw egocentric footage, what’s considered usable may be only five minutes or less, he argued.

Repeating the ChatGPT playbook

Sources Tech in Asia spoke to described the demand for egocentric data as broad and still immature. Robotics labs want to train for domestic chores such as folding clothes and cleaning dishes as well as tasks within industrial and commercial settings.

The point is not simply volume. Buyers want diversity in the data across people, environments, geographies, lighting conditions, tools and the kinds of mistakes that are made.

Keechin Goh, co-founder and CEO of Singapore-based vision AI startup Datature, said many robotics companies are trying to repeat the broad playbook that worked for ChatGPT. That means collect enough relevant data, pre-train a model, then refine it later with more specific feedback.

That doesn’t mean robotics labs are starting from zero, as open egocentric datasets already exist. Companies like Nvidia have also released foundation models for labs to tinker with for free, but Goh said these resources have not removed the need for proprietary data collection.

Public datasets are still limited, he added, especially when compared to data available to AI companies in the early stages of developing foundation models.

“Until then, people are collecting whatever they can,” Goh noted.

MyRobin’s Kumar said his company began receiving inbound interest around late 2025. He initially ignored the requests, but after several companies reached out, he realised that robot data collection could become a new business line.

To work with MyRobin, the first step for a client is to define the type of task the robots need to be trained on and the number of hours required. They then send cameras to the company or a field partner.

Workers are onboarded and instructed to keep their hands and objects in frame, avoid faces and turn off cameras during breaks. Once recording is finished, the SD cards or video uploads are sent to a dashboard, where MyRobin checks for lighting, hand visibility, idle time, and other defects. Bad footage may be rejected and required to be recorded again.

That said, data collection is only the first layer. The footage must be trimmed, labelled, anonymised and converted into formats that robotics labs can use.

Kumar acknowledged that egocentric data collection hasn’t generated significant revenue yet for MyRobin, but he expects the business to evolve as clients ask for more than just first-person video and start sending in gloves and other types of devices to capture different stimuli.

The problem of easy money

It isn’t expensive to get workers to start recording videos, either.

It often requires only a low-end Android phone and a headband to hold it, said Nikita Sadhnani, founding member of San Francisco-based BergLabs, which works on video quality control and action labeling.

But some in the industry are concerned that video captured in factories isn’t the goldmine that some make it out to be.

Michael Cho, co-founder and CEO of Singapore-based Frodobots, which builds specialty cameras to capture egocentric videos, said robotics needs real-world diversity.

Workers recording themselves repeating the same action on an assembly line has limited value. After a few days, there may be little new information left to collect, and the company often moves on quickly once this happens.

“That kind of setting doesn’t give you much entropy,” he added.

In the context of machine learning, entropy simply means useful variation. Things may change constantly inside a restaurant or on a sidewalk. Meanwhile, a factory is designed to do the opposite by standardising almost everything and removing surprises as much as it can.

The overall value of the egocentric video market is still hard to gauge because the industry is still young, but Humyn Labs’ Agarwal said his company has about US$5.5 million in signed contracts and has a contracted pipeline of US$40 million to US$50 million.

India and South-east Asia, specifically, have become important collection hubs because workers can be recruited at lower cost and at larger scale. But that advantage comes with a catch: It may make the region a source of cheap raw materials and not necessarily a winner in the next stage of robotics.

Sayantan Sarkar, an investor at Stellaris Venture Partners, said the raw collection business has little technological moat. The hardware is relatively easy to procure or assemble, and the business is mostly about getting cameras into the hands of enough people across enough settings.

The real prize, he noted, is having direct access to frontier labs such as OpenAI or Physical Intelligence and being able to sell them high-quality data with annotations that are “worth far more than raw footage”.

Agarwal agreed. He said unprocessed footage coming from a single, sight-only camera sits at the low end of the quality spectrum.

Setups that capture more than a flat video feed are more important. Multiple synchronised cameras can help show where a hand or object is in 3D space, while touch and motion sensors can show how much force was used or whether an object slipped during handling.

Agarwal said the price of raw egocentric data has already fallen over the past year, from around US$6 to US$10 per hour of footage to roughly US$2 to US$3 today. The reason, he added, is that more suppliers have realised they can recruit lower-cost workers, record large amounts of video and sell it to labs while demand is still hot.

“If we stay stuck at raw collection, we become a useless commodity in nine to 12 months,” he argued.

Buyers are also becoming more cautious over concerns about duplicate datasets being resold despite exclusivity agreements, said BergLabs’ Sadhnani.

Some labs are also bringing collection and processing in-house, while others are relying on preferred vendors that can prove quality and scale.

The future won’t be egocentric

For now, however, labs are still buying imperfect data. In the absence of high-quality multimodal data, they buy what exists, discount for quality and process it themselves, Agarwal said.

The arrangement for egocentric videos also raises an obvious question for workers: Are they being paid to help train machines that may one day replace them?

Kumar of MyRobin said he’s not aware of workers directly raising that concern, though he assumes some may have asked. His view is that some tasks will eventually be automated, while new human-in-the-loop roles will emerge, though he didn’t elaborate further.

Several sources agreed that the next wave of this robotic data collection boom will involve workers wearing not only cameras but also gloves or other input devices that capture interaction forces, tactile information and hand motion.

That shift is already beginning, with buyers asking about gloves, wrist cameras, depth sensors or even arm-tracking systems to be included in the collection process. But these are more expensive and harder to scale than a phone on a headband.

“The switching back and forth (between vision and touch) is the most important thing. Our brains are very good at doing that,” Hand Plus Robotics’ Causo noted.

The likely future is not the disappearance of egocentric video but rather its demotion. Instead, specialised and well-labelled interaction data will become the premium product.

According to Causo, “The reckoning is not going to happen quickly. It will happen six or 12 months down the line.” TECH IN ASIA