Picture a room full of humanoid robots doing laundry, opening fridges and rearranging a sofa. Now picture a human operator steering each move by remote control. That’s not a sci-fi set. It’s South Korea’s new humanoid data factory, and it opened on November 3 in the city of Daejeon.
Why would anyone build robots that need a person to drive them? Because that’s how you teach them. And right now, teaching is the hardest part of the whole humanoid business.
So what opened in Daejeon?
The facility is called the AI Humanoid Data Factory. Two Korean research institutes built it together: the Korea Institute of Machinery and Materials (KIMM) and the Electronics and Telecommunications Research Institute (ETRI). It sits at KIMM’s headquarters, with data collection spaces on two floors of one building. Korea’s Ministry of Science and ICT announced the opening and backs the project.
KIMM calls it “the country’s first dedicated data production facility for bipedal humanoid robots.” Bipedal just means two legs, like us. So this is about full human-shaped robots, not robot arms bolted to a table.
For now, it runs 12 humanoids. That’s a modest number, honestly. But KIMM plans to add more robots and more types of tasks over time. It hasn’t given a target figure yet.
The project also ties into a national program called the K-Moonshot AI Humanoid Mission. In other words, this isn’t a one-off lab experiment. It’s part of a bigger plan to build a homegrown humanoid industry.
How a humanoid data factory actually works
Here’s the basic loop. First, human operators control the robots remotely. While they do, the system records every movement along with details about the task. That recording becomes training data.
Next, researchers use that data to train AI models. The goal is a model that understands what a person asks and what’s around the robot, so it can eventually do the job alone. Then they load the trained model back onto real robots and test it. Data from those tests feeds into the next round of training. Repeat, and the robot gets a little better each cycle.
What the robots practice
The tasks are a mix of home chores and factory work. On the home side, robots take items out of a refrigerator, handle laundry and arrange a sofa. On the industrial side, they press buttons and work levers inside a vehicle. The data covers arm and leg movements, plus full-body tasks.
At the opening, Korea’s own humanoid, KAIROS, showed off precise object handling and copied full-body motions. KIMM also wants the models to work across different robot bodies, including a shared design called the K-Humanoid standard platform. That’s smart, because nobody wants to retrain from scratch for every new robot.
Why robots are starving for data
Here’s the way we think about it. Chatbots got clever by reading huge amounts of text. But there’s no giant online library of “how to grab a slippery cup” written in motor commands. Robots have to collect that experience in the physical world, one move at a time.
The ministry made the same point in its own words. Humanoids that work on their own in homes or factories have to learn real movements, like sensing what’s around them and gripping objects. That takes a lot of varied task data. The ministry even called data factories “core infrastructure that determines competitiveness in the humanoid industry.”
KIMM President Seog-Hyeon Ryu agreed: “Securing large volumes of high-quality robot task data is essential to national technological competitiveness.” He also called the facility “an important starting point for building Korea’s own robot AI training data.”
This is also why robot “brain” companies are drawing so much money. We wrote recently about Physical Intelligence hitting a $20 billion valuation and about renting a robot’s smarts for $3 an hour. That’s because whoever owns the best training data has a real head start.
The China factor
So why now? Look next door. According to Interact Analysis, cited by Seoul Economic Daily, China had at least 90 humanoid data collection and training centers operating or planned as of the end of April. At least 64 were already running. And 13 of them reportedly had more than 100 robots each.
Shanghai’s Pudong New Area alone reportedly gathered over 50,000 hours of multimodal data from real robots. Multimodal means several kinds of data at once, like video plus motion readings. So next to that, 12 robots in Daejeon looks small.
Still, Korea’s goal isn’t to win on headcount tomorrow. Officials say they want to cut the country’s reliance on overseas data and AI models. That’s a sovereignty play as much as a tech one. If your robots learn only from someone else’s data, you don’t fully control what they know.
We’ve seen Chinese firms move fast on the factory floor too, like when a Xiaomi humanoid worked a full 8-hour shift. Korea clearly doesn’t want to sit that race out.
What a humanoid data factory means for smaller tech countries
The most interesting part might be the plan to share. Starting at the end of 2027, KIMM will gradually open the facility to Korean companies, universities and research institutes. That turns the data factory into shared public infrastructure, a bit like a national lab or a supercomputer that startups can book time on.
For a country like Armenia, that model is worth studying. We’ll probably never match China’s robot count. But a small, well-run humanoid data factory, open to local startups and universities, could give a whole ecosystem a shot at building robot AI instead of only buying it. And since the data comes from teleoperation, you need skilled people as much as fancy hardware.
So here’s our open question. If data is the new bottleneck for robots, should smaller countries build shared data hubs now, before the market decides for them? Korea just answered yes. We’ll be watching to see if 12 robots turn into a lot more.
Sources: Seoul Economic Daily, KIMM release via EurekAlert and The AI Insider.