A robot data factory is exactly what it sounds like, a building full of robots and human operators where the only product is training data. No widgets come off the line. Nothing ships in a box. People drive robots through tasks all day, every motion gets recorded, and the recordings are the inventory of the robot data factory.
Six days ago, South Korea opened its first one, and it’s worth paying attention to because of what it says about where this industry is headed.
The facility is called the AI Humanoid Data Factory. It sits on two floors at the Korea Institute of Machinery and Materials in Daejeon, about a thousand square meters, built with the national electronics research institute and backed by the science ministry. Twelve humanoids run there now, a mix of machines including Korea’s own KAIROS, with plans to grow to twenty five. Operators teleoperate them through household chores like pulling items out of a refrigerator, handling laundry, and rearranging furniture, plus industrial work like operating buttons and levers inside a vehicle. Every joint angle and camera frame gets captured and stored. Starting in late 2027 the whole thing opens up to outside companies, universities, and labs.
This AI Humanoid Data Factory is a prime example of how a robot data factory functions, focusing on gathering quality robotic data.
Such a robot data factory represents a significant investment in national technological capabilities.
A robot data factory exists to solve the challenges of data scarcity in robotics, ensuring a sustained supply of useful data.
The model training process in a robot data factory relies heavily on the quality of the collected data.
Only a well-structured robot data factory can ensure consistent quality in the task data generated.
The advancements from a robot data factory are pivotal for the future of robotic applications.
The KIMM president’s line at the ribbon cutting is the tell. He called large volumes of high quality robot task data essential to national technological competitiveness. Read that again. A head of a national research institute just described recorded robot demonstrations the way governments used to talk about steel and semiconductors. Korea has put roughly 1.7 billion dollars behind its humanoid industry through 2030, and the first big thing they built with it wasn’t a robot plant. It was a data plant.
If you’ve read this blog before you know the root problem. There is no internet of physical skill. Language models got to eat trillions of words that humans had already written for free. A robot learning to fold laundry gets nothing, because nobody ever wrote down the motions. The data has to be manufactured, and the most direct way to manufacture it is teleoperation. A trained person drives the actual robot through the actual task, and the recording becomes ground truth a model can learn from.
For a few years that happened in labs, a rig here, a grad student there. What changed is scale. The industry figured out that demonstration data behaves like any other feedstock. More of it, at higher quality, produces predictably better robots. And the moment something becomes feedstock, somebody industrializes it.
That’s what’s happening everywhere at once right now. China has announced more than forty state backed data collection centers, with roughly two dozen already operating. Shanghai’s Pudong district alone has reportedly gathered over fifty thousand hours of multimodal robot data. JD.com and a partner city set a target of ten million hours of robot training data inside two years, with plans for a collection workforce in the hundreds of thousands. UBTech has sold about eighty million dollars worth of humanoids to data collection centers, meaning robots are being bought specifically to be teleoperated for data. In the US, Tesla hires data collection operators at up to 48 dollars an hour to generate demonstrations for Optimus, and Apptronik has described running fleets of teleoperated robots at a large facility and learning that building more robots was the fastest way to accelerate data collection. At RoboBusiness this month, industry leads flat out named data collection as one of the biggest obstacles standing between humanoids and real commercial work.
Factories, farms, foundries, pick your metaphor. The point is the same. Teleoperation data collection has stopped being a lab activity and become infrastructure.
Here’s the thing the headlines miss, and it’s the difference between a data factory and an expensive room full of robots.
Understanding the operations of a robot data factory is crucial for stakeholders in the robotics field.
Ultimately, the success of any autonomous system hinges on the efficiency of its robot data factory.
Hours are not the metric. Usable episodes are. A demonstration only counts if it was done right, by an operator who knows the robot’s dynamics, captured with synchronized streams, and passed through quality control before it touches a training set. A sloppy demonstration isn’t a wasted hour. It’s worse than that, because it teaches the model something wrong, and you may not find out until the robot fumbles on somebody’s floor. Teams in this space regularly see rejection rates north of thirty percent when collection is treated as a volume game instead of a craft.
Which means the scarce resource inside every one of these buildings isn’t the robots. It’s the people and the pipeline. Skilled operators are the machinists of this new kind of factory. The QA layer is the quality department. And simulation is the force multiplier that takes a thousand clean demonstrations and turns them into the hundred thousand variations a model actually needs, different lighting, different part positions, the edge case that shows up once a month in real life but ten thousand times tonight in sim.
One of the write ups on the Korean facility put it plainly. You need skilled people as much as fancy hardware. That sentence is the whole industry right now.
This is the layer FS Studio operates in, and not theoretically. We run teleoperation data collection at production scale for one of the biggest names in AI, staffing trained operators on humanoid platforms daily and running the pipeline that turns thousands of raw demonstrations into clean, model ready training data, backed by the synthetic data generation that fills in everything the real sessions can’t cover. The lessons the new data factories are about to learn, about operator skill, capture quality, and what separates data a model learns from and data that quietly poisons a run, are lessons we’ve already paid for.
So when a national lab cuts a ribbon on a building whose only product is robot training data, we don’t see a curiosity. We see validation. Countries are now doing at a policy level what the smartest robotics teams started doing two years ago, treating demonstration data as manufactured goods with a supply chain, a workforce, and a quality standard.
The robots get the headlines. The buildings teaching them are the real story. And if your team is building toward autonomy and the data layer is the thing slowing you down, that is exactly the conversation we want to have.
By Bobby Carlton On July 28, 2026, the FCC added foreign-produced advanced robotic devices to…
Embodied Reasoning Is Manufactured, Not Discovered By Bobby Carlton Embodied reasoning is the term you're…
By Bobby Carlton Where Humanoid Robot Reliability Actually Comes From Humanoid robot reliability crossed a…
Why Teleoperation Data Can't Be Faked. By Bobby Carlton Teleoperation data is the quiet force…
What Better Robot Simulation Actually Means for Your Floor By Bobby Carlton Here's something that…
Good chance you’re hearing a lot of conversation around a technology called Gaussian splatting (or…