In the kitchens and living rooms of Indonesia and beyond, a quiet labor is unfolding: thousands of gig workers strap cameras to their heads and film the ancient rhythms of domestic life — washing, folding, sweeping — so that machines might one day learn to do the same. For a few dollars an hour, workers like twenty-three-year-old Mohamad Dunggio from Sulawesi are becoming the unseen teachers of a robotic future, their ordinary gestures transformed into datasets that cross oceans and corporate ledgers. The arrangement raises old questions in new forms — about who benefits from labor, who owns a
Indonesian gig workers film chores to train AI robots for pittance
Some videos rejected, some earnings charged a dollar fee to withdraw
Mark
So Dunggio spent four hours filming himself wash dishes and made between three and ten dollars. That's the entire transaction?
Mimi
That's one transaction. He's done multiple videos for different companies—cooking, cleaning, folding clothes. Some get rejected. And when he tries to withdraw his money, there are fees.
Luke
Wait—rejected by whom? On what basis? The source doesn't say. We know he made videos, we know some were rejected, but we don't know the criteria or who decided.
Mimi
Fair point. But the pattern is clear: low pay, unpredictable acceptance, friction in getting paid out. That's the worker experience.
Mark
Why would companies want egocentric video specifically? Why not just film robots doing the tasks?
Mimi
Because robots doing the tasks are expensive to set up and operate. Human video is cheaper, faster, and captures way more variation—different kitchens, different techniques, different hands.
Luke
But Kulić said robot-collected data might be more relevant. So there's a trade-off between quality and cost. The source doesn't tell us which one matters more for actually building working robots.
Mark
And do these robots actually work yet?
Mimi
Not reliably, according to Kulić. There's a huge difference between a video of someone folding a shirt and a robot that can fold shirts every day without breaking.
Luke
So we're collecting millions of hours of video to train robots that don't exist yet and might not be economically viable when they do. And the people providing the training data have no idea what happens to their videos.
Mimi
Exactly. Privacy, data ownership, awareness—none of it is settled.
Mark
How many people are doing this?
Mimi
Appen alone works with up to five thousand participants across multiple countries, collecting sixty thousand hours per month. But that's just one company.
Luke
The source doesn't say how many companies are doing this or how many workers globally. Five thousand is a floor, not a ceiling.
Le Pouls
- A young Indonesian man earns as little as three dollars an hour filming himself wash dishes, only to find a withdrawal fee waiting when he tries to collect ten dollars in earnings.
- Companies are racing to harvest tens of thousands of hours of first-person household video each month, betting that human recordings are cheaper and richer than anything a robot could film itself.
- Workers often lack clear information about where their footage goes, who will own it if it is resold, and what privacy risks come with filming the interior of their own homes.
- Robotics experts caution that the gap between a well-filmed chore and a reliable household robot remains enormous — manufacturing costs, maintenance, and real-world durability are all unsolved problems.
- The industry is quietly expanding, sustained by wage differentials that make three to ten dollars an hour feel significant in some economies while the data collected may eventually power machines worth thousands.
In the kitchens and living rooms of Indonesia and beyond, a quiet labor is unfolding: thousands of gig workers strap cameras to their heads and film the ancient rhythms of domestic life — washing, folding, sweeping — so that machines might one day learn to do the same. For a few dollars an hour, workers like twenty-three-year-old Mohamad Dunggio from Sulawesi are becoming the unseen teachers of a robotic future, their ordinary gestures transformed into datasets that cross oceans and corporate ledgers. The arrangement raises old questions in new forms — about who benefits from labor, who owns a moment captured inside one's own home, and how far the distance truly is between a filmed shirt and a robot that can fold one.
Mohamad Dunggio, twenty-three, strapped a phone to his head in April and spent four hours recording himself washing dishes. The footage would become training data for robots. Companies based in Argentina, India, and the United States paid him between three and ten dollars per hour — and when he tried to withdraw his accumulated ten dollars, one company charged him a dollar fee to access it.
Dunggio is one of thousands of gig workers now filming the mundane choreography of domestic life — dishwashing, cooking, folding clothes, sweeping — to feed machine learning systems. The recordings are called egocentric data: first-person footage, typically from a head- or chest-mounted camera, that keeps the performer's hands visible and centered. This angle matters because robots must learn to navigate the specific, unpredictable conditions of real homes, where no two kitchens are identical and no two people fold a shirt the same way.
Robotics professor Dana Kulić of Monash University explained the appeal: human recordings are cheaper, faster to collect in volume, and capture far greater real-world variation than footage gathered by remotely operated robots. Appen, an Australian data aggregator, collects roughly sixty thousand hours of such video per month from up to five thousand participants worldwide, with China its largest robotics-data market and North America its emerging focus.
Yet the distance between collecting video and building reliable robots remains vast. Kulić stressed that filming a shirt being folded is fundamentally different from engineering a machine that can fold shirts day after day without failure. Manufacturing costs, maintenance, and long-term reliability remain open questions, and practical household robots remain, in her view, a distant prospect.
For Dunggio, the work brought one small, human reward: he learned to cook by filming himself do it. But larger questions remain unresolved — how aware are these workers of how their data will be used, what privacy protections apply when someone films inside their own home, and who owns the recordings if they are sold to third parties. The industry grows quietly, powered by workers for whom a few dollars an hour represents real income, even as the data they create may one day animate machines worth thousands.
Mohamad Dunggio, twenty-three years old, strapped a phone to his head in April and spent four hours recording himself washing dishes. The video would become training data for robots. For his effort, companies based in Argentina, India, and the United States paid him between three and ten dollars per hour.
Dunggio is one of thousands of gig workers across multiple countries now filming the mundane choreography of household life—dishwashing, cleaning, cooking, bed-making, ironing, folding clothes, sweeping—to feed machine learning systems. He lives in Gorontalo on Sulawesi, in Indonesia, and has turned this work into a side income. Over the months, he submitted videos of various tasks to different companies. Some were accepted. Some were rejected. When he tried to withdraw his earnings, which accumulated to around ten dollars, one company charged him a dollar fee just to access the money he had made.
The videos these workers create are called egocentric data—recordings made from the perspective of the person performing the task, typically with a camera mounted on the head or chest so that the performer's hands remain visible and centered in the frame. This first-person angle matters because robots need to learn how to observe and respond to the specific conditions they will encounter in actual homes, where no two kitchens are quite the same and no two people fold clothes identically.
Dana Kulić, a robotics professor at Monash University, explained the logic behind this approach. Robots could theoretically learn from videos recorded by other robots, remotely controlled by humans, but that method is expensive and time-consuming. Human video recordings are cheaper to collect, faster to gather in volume, and they capture a far wider range of real-world scenarios and variations. The trade-off is that the data comes from people like Dunggio, working for minimal pay, often without full clarity about how their recordings will be used or where they might end up.
Sergio Bruccoleri, vice president of delivery at Appen, an Australian company that aggregates this kind of training data, said his firm collects roughly sixty thousand hours of video per month from as many as five thousand participants scattered across different countries. China remains Appen's largest market for robotics data, followed by the United States. The company is now shifting its focus toward data collected in the United States and Canada, betting that these are the regions where household robots will first find commercial buyers.
Yet the gap between collecting video and building reliable robots remains vast. Kulić emphasized that demonstrating how to fold a shirt on camera is fundamentally different from engineering a robot that can fold shirts reliably, day after day, without breaking down. Questions linger about the actual cost of manufacturing these machines, maintaining them over time, and repairing them when they fail. Despite years of technological advancement, practical humanoid robots capable of handling the full range of household chores remain, in her assessment, a distant prospect.
For Dunggio, the work has had at least one unexpected benefit: he began cooking on his own, learning through the process of filming himself prepare meals. But the broader picture raises uncomfortable questions that neither the companies nor the workers have fully resolved. How aware are these gig workers of exactly how their data will be used? What privacy protections exist when someone is filming inside their own home? And who owns the recordings if they are sold or transferred to third parties? These questions hang over an industry that is growing quietly, powered by workers in countries where three to ten dollars per hour can feel like meaningful income, even as the data they create may eventually power machines worth thousands.
Citations marquantes
Robots need to learn to observe and respond to their environment, as conditions in homes differ— Dana Kulić, Monash University robotics professor
There is a significant difference between demonstrating clothes folding and a robot working reliably over a long period— Dana Kulić
Envie de l'histoire complète ?
Lire l'original sur ua.news ↗
Technology