Executive Summary

Key Numbers

The robot count and the budget are the size KIMM announced. How small a dataset built by teleoperation is next to other fields is an estimate from outside, and the opening date is a schedule that has not happened yet.

Sources: ZDNet Korea (2026-11-03), KIMM opening release (2026-11-04), The Robot Report (2026-07-19)

The body behind the facility is the Global TOP Strategic Research Group for Self-Growing AI Humanoids, which KIMM leads. Chan Hun Park directs it, and ETRI built the site with them. It is one of ten groups the Ministry of Science and ICT and the National Research Council of Science and Technology set up in May 2025 with ₩610 billion between them over five years, and this group's share is ₩21 billion a year, ₩105 billion across the five. The term runs from May 2025 to April 2030. ETRI and the Korea Institute of Industrial Technology take part alongside KIMM, and nine Korean universities including KAIST and GIST plus five research institutes abroad are attached. On the industry side the member list carries A-Robot, Lion Robotics and LG Electronics, and the Korea Robot Industry Association joins as a supporting organization.

What the place is for is exactly what the name says. It does not build robots and it does not train models; it manufactures the data that will teach a robot how to do a job. Securing everything from arm and leg movement to work that uses the whole body, all in one place, is how KIMM describes the use. Seog-Hyeon Ryu, president of KIMM, said at the opening that "securing large volumes of high-quality robot task data is essential to national technological competitiveness."

One piece of background belongs here. When the research group launched, its plan carried a second item paired with the data factory. The group said it would build an open AI platform that shares the data it gathers with outside researchers and takes in AI models built elsewhere, and that the pair of them would make a K-humanoid open ecosystem. What opened this week is the half that makes the data.

A person is inside the collection method. An operator drives the robot by remote control to make it do a job, and the movement and task information thrown off along the way get recorded and used for AI training. A second method runs alongside it, in which the robot copies a person's own movements. Either way, the person never steps out. Producing one record takes one person working for its whole length.

The two methods do not leave the same record behind. Teleoperation has a person moving the robot's body directly, so the values that come out of the robot's joints are what gets stored. Motion imitation has the person do the job with their own body while the robot carries that movement over into its own joints, so the work of matching a human body to a robot body goes into the record too. What KAIROS showed at the opening was these two as well, precise object manipulation and full-body motion imitation.

The task list at the opening sits closer to the inside of a house than to a factory floor. Taking items out of a refrigerator, handling laundry and arranging a sofa went on as the everyday work, and operating the buttons and levers inside a car was attached as the industrial work. Housework came first for a reason other than being easy. Pulling a refrigerator door open takes bracing on the legs, twisting at the waist and changing the angle of the wrist all at once, and laundry and other objects that hold no fixed shape never let the same motion be used twice. For pulling out data where the whole body works instead of an arm here and a leg there, nothing beats everyday chores.

Twelve robots stand on this floor today. Chinese and Korean machines are mixed together, and KAIROS, which KIMM is developing itself, is among them. Not standardizing on one model is a deliberate choice. Chan Hun Park, the group's director, said a data factory ought to gather its data from a range of robots in the first place. KIMM set out a plan to grow to twenty-five by 2030 and to widen the robots and the tasks it collects step by step.

The site's use does not stop at piling up records either. KIMM describes a loop that runs inside the building: the gathered data trains an AI, the trained AI goes back onto a physical robot so its performance can be checked, and the records that come out of that go back into training. The place that makes the data and the place that puts a model built from that data onto a real robot sit in the same building.

Having a person drive a robot to produce records is very nearly the standard way humanoids are trained right now. So an argument against the method itself, written from inside the industry, is worth reading. Nikita Rudin, who leads the Zurich humanoid AI startup Flexion Robotics, made one in the American robotics trade outlet The Robot Report in July, on how to avoid what he calls the teleoperation trap. Rudin worked on large-scale reinforcement learning and robot simulation at NVIDIA before founding the company, and what he aims at is not any one institution but where the industry is putting its money.

Rudin carries one number. Teleoperation datasets today are over 100,000 times smaller than what is used to train current language and vision models. Try to close that gap with people, though, and the ceiling on speed lands on the human side. Every variation needs a fresh demonstration, and somebody has to perform it, so the data can only grow as fast as human labor allows. On top of that, an operator cannot feel what the hand is touching and cannot judge depth reliably, so they move slowly and overcorrect, and that habit gets learned along with the task.

This warning cannot be laid straight over the Data Factory. Rudin himself allows that for narrow, repetitive tasks in controlled environments, collecting this way is fair enough, and a record left by a real robot touching real objects holds physical fact that no simulation can supply in its place. The ceiling he pointed at stays, though. Whether the robots go from twelve to twenty-five, records come out no faster than the number of operators and the hours they work. So whether this facility pays off turns less on how many robots stand in it than on how many times one record can be reused.

The opening plan KIMM gave has dates attached. From the end of 2027 the facility opens in stages so that companies, universities and research institutes can use it together. What each side gets is split, too. Universities and research institutes get an environment for studying and evaluating AI models, and companies get a chance to collect data and run demonstrations. President Ryu said KIMM would speed up work on KAIROS through industry, academic and institutional cooperation and open the Data Factory from the end of 2027, a phase at a time, to strengthen Korea's physical AI industry. Kim Sung-soo, who heads the R&D policy office at the Ministry of Science and ICT, said the ministry would widen the base for producing and using data that industry, academia and research institutes can share.

That is as far as anything is settled. Nowhere in the opening announcement or the coverage around it does it say what format the data gets stored in, on what terms it is handed over, whether commercial use is allowed, or who qualifies to apply. More than a year sits empty between the day the facility went up and the day a researcher outside it receives the data. The robots keep running through that gap.

Leaving that cell blank is not a Korean peculiarity. At the International Organization for Standardization, WG 16 under ISO/TC 299 is writing a standard on humanoid robot datasets. The document number is ISO 26264-1 and the title is "Humanoid robot datasets — Part 1: General requirements." It frames the life of a dataset from the moment it is planned to the moment it is disposed of, and sets quality, security and privacy as common requirements, covering bipeds and wheeled machines alike across industrial, service and medical work. It was registered as a committee draft at the end of June 2026 and the comment period closed at the end of August. It is the first international standardization effort aimed at humanoid robot datasets alone.

The bottleneck this standards work points at is not only that there is too little data. It is that collection costs a lot, that what gets gathered stays locked inside each organization, and that evaluation criteria differ from place to place, so the data never accumulates. Record the same task in two different places, and if body, action, scene and execution trace are written down under different names, the two records do not add up. That is why equipment keeps arriving while experience does not accumulate. The Data Factory's choice not to standardize on one robot model buys diversity and takes this problem head-on in the same move.

That diagnosis is written out in a paper the researchers drafting ISO 26264-1 published separately in June. In it they argue that whether humanoid data can be used again later rests on physical coherence. Reusing streams that arrived along several channels means each record has to show on its own what the clocks were set to, where the coordinate frames sit, when and how the sensors were calibrated, and what units lengths and forces were written in. Even with no format fixed yet, in other words, what has to be fixed already has names.

So what actually gets decided in the stretch running to the end of 2027 is not how busy the facility is but what shape the record takes. Which cells get filled now effectively settles who can use this data later, because an entry nobody wrote down does not get created retroactively.

Pebblous has split the data used in robot learning into five layers before and ranked them on quality, diversity, reusability and physical fidelity (What Do You Feed a Robot?). Data gathered on real robots came out highest on physical fidelity and lowest on reusability. What this Data Factory produces is exactly that layer. It is the most expensive and the closest to the real thing, and left alone it travels the least.

There is also a piece that reads the ISO 26264-1 draft on its own (Robots Are Multiplying. Why Isn't Their Experience?). The question that draft raises lies directly over this opening. Add equipment and the data grows, but experience does not accumulate by itself. What makes it accumulate is the record's format and the terms under which it changes hands.

A data practitioner has three things to track here.

• When the record format gets published. Only a published format lets anyone outside produce data in the same shape and merge it in.
• The terms of use. How far commercial use extends and whether redistribution is allowed decide whether a startup can touch this material at all.
• The distance to the international standard. Once ISO 26264-1 sets, lining up with it means reworking records already piled up.

All three are easier to answer from the data side than from the robot-building side, because that is where the people sit who know which values land where and on what cycle. The state has now spent public money to confirm one thing. The next contest in physical AI is not the model but the record a person makes by driving a robot, and what that record is worth is settled by the form it is kept in, not by the moment it is collected.

Thanks for reading this far. The original coverage of the opening is at ZDNet Korea and AI Times Korea, the English release is on EurekAlert, and the research group's original plan is in the KIMM press release. If your team is already collecting robot task data against a fixed schema, we would like to hear which fields you put in and which you left out.

News Coverage

• 1.ZDNet Korea. (2026). "국내 첫 휴머노이드 데이터 팩토리 구축…'서비스는 2027년 말'" [Korea's First Humanoid Data Factory Built, Service at the End of 2027]. ZDNet Korea.
• 2.AI Times Korea. (2026). "기계연·ETRI, 국내 첫 AI 휴머노이드 데이터 팩토리 개소" [KIMM and ETRI Open Korea's First AI Humanoid Data Factory]. AI Times Korea.
• 3.AI Times. (2026). "정부 주도 'AI 휴머노이드 데이터 팩토리' 문 열어" [Government-Led AI Humanoid Data Factory Opens]. AI Times.
• 4.Robot Newspaper. (2026). "휴머노이드 로봇 학습, '텔레오퍼레이션 함정' 벗어나야" [Humanoid Robot Learning Has to Escape the Teleoperation Trap]. Robot Newspaper.

Official Announcements and Standards

• 5.National Research Council of Science & Technology. (2026). "Korea opens its first AI humanoid data factory." EurekAlert.
• 6.Korea Institute of Machinery and Materials. (2025). "산업에서 일상까지, 인간과 함께 할 휴머노이드 동반자 실현" [From Industry to Daily Life, Realizing Humanoid Companions]. KIMM NEWS.
• 7.Korea Institute of Machinery and Materials. (2025). "글로벌 TOP 전략연구단 출범, 한국기계연구원 AI 휴머노이드·질소 자원화 총괄" [Global TOP Strategic Research Groups Launch, KIMM Leading AI Humanoids and Nitrogen Resource Recovery]. KIMM NEWS.
• 8.ISO. (2026). "ISO/CD 26264-1 Humanoid robot datasets — Part 1: General requirements." ISO/TC 299/WG 16.

Commentary and Papers

• 9.Rudin, N. (2026). "How to avoid the teleoperation trap in robotics development." The Robot Report.
• 10.Liu, S. et al. (2026). "Data Standards for Humanoid Robotics: The Missing Infrastructure for Physical AI." arXiv:2606.19769.