Executive Summary

Key Numbers

Source: Seoul city and SNU College of Engineering agreement announcement (October 30, 2026), ZDNet Korea reporting (November 1, 2026)

On October 30, Seoul National University's College of Engineering and the Seoul city government signed a business agreement on physical AI at the city's Seosomun Building 2. Kim Young-oh, dean of the College of Engineering, signed for the university, and Kim Hyun-woo, head of the city's economic affairs office, signed for Seoul. Seven people attended the ceremony, among them Cho Kyu-jin, director of the SNU Robotics Institute. The agreement runs three years and can be extended if both sides agree.

Space is the substance of the agreement. Seoul will build a 612-square-meter Physical AX Center at the main building of Seoul AI Hub in Yangjae-dong, Seocho-gu, and as the first phase will fit out a 338-square-meter H-PAIR everyday-support living lab on the fourth floor. The number 338 is the size of the first room to open, not of the whole center. Let that distinction slip and the center reads as roughly half the size it is.

Inside the living lab go a living room, a kitchen and a bedroom, laid out at full scale along the paths people actually walk. Lighting can be altered, objects can be pushed out of place and fresh obstacles can be set down, so the differences that run from one home to the next are produced deliberately. The space is to be finished by the first half of 2027, and verification of care robots, mobility-assist robots and other living robots begins in the second half. Half a year separates the day the space opens from the day the robots arrive.

The announcement names five kinds of robot for testing: strength-assist robots, mobility-assist robots for people who have trouble getting around, wearable robots, care robots and serving robots. There are three axes of measurement as well: performance, safety and user acceptance. The first two look at what the robot does; the third asks whether the person using it will accept it.

H-PAIR is an interdisciplinary body built around the SNU Robotics Institute, drawing in 16 departments and graduate programs and 41 full-time professors. It was selected as a national research laboratory in 2026 and receives ₩95 billion in state funding across the ten years from 2026 to 2035. Some outlets have put the figure at up to ₩100 billion. Its research runs along three branches, everyday support, body-augmenting wearables and intra-body robots, and this living lab takes on the first of them. Wearable robots also sit among the five kinds of robot listed for testing, so the branches do not divide cleanly; the announcement sets the two lists down separately and does not say how they relate.

A division of labor sits behind the split between Gwanak and Yangjae. The H-PAIR laboratory on the Gwanak campus draws on the university's research resources and holds the core of fundamental technology work, while the Physical AX Center in Yangjae becomes the forward base that connects what comes out of there to companies and pushes it through verification, field demonstration and commercialization. That is how the agreement announcement explains it. Seoul takes the space and the cooperation structure, SNU the technical verification and startup incubation. The reason for a separate site in Yangjae lies in that division.

The route companies take in is organized under the name Physical AX Consortium. Care centers, senior housing and welfare centers under the Seoul city government put forward the problems they run into on site, and companies develop AI and robot technology against those problems. What they build goes through technical advice and verification at the H-PAIR living lab, then returns to the field as a pilot. One consortium is to run as a trial in 2027, and from then on about two projects are added each year.

What the laboratory will hand over to companies also has a list. MORPHI bundles variable-stiffness and variable-shape structures with contact-safe electronic skin and a tactile sensor design library. MOSAIC is a distributed-intelligence software development kit covering intention, emotion and situational awareness. Standard test protocols and evaluation kits follow, and all of it opens in stages. On top of that, a startup seed fund run with SNU Holdings is to find five promising startup teams by 2028, specialist training runs four times a year for 10 to 20 people a session, and there are undergraduate field placements and an annual hackathon.

Picking everyday support as the first phase out of the three research branches is a fact in the announcement. Why that choice leads to a data problem is something the announcement does not say.

A robot running in a factory can have the environment bent to suit it. Lines go down on the floor, lighting is held steady, and the process is arranged so that parts always arrive in the same place. A home allows none of that. Living room light changes by the hour, the things on the table are not where they were yesterday, and something is lying on the floor that was not there before. However large the model grows, without records of scenes like these the robot starts over, lost, in somebody else's living room.

The most striking part of the living lab's design is that lighting, the placement of objects and obstacles were all made changeable. If the point were to copy one house exactly, that capability would be unnecessary. Making it changeable reads as an intention to produce several sets of records under different conditions within the same space. What is being built is not one room in a laboratory but a stage where the variables themselves can be moved.

The direction demand flows in points the same way. Rather than a research project being set first and the field then searched for, a project arises when a care center or a welfare center puts up something it is stuck on. Seoul describes the agreement as the working out of the "physical AI leading city, Seoul" vision it announced at AI SEOUL 2026 in January, and mentions alongside it a Seoul physical AI belt linking the Yangjae AI cluster to the Suseo robot cluster.

A day before the agreement, on October 29, the Ministry of Science and ICT held a meeting of industry, academic and research experts at the Korea Chamber of Commerce and Industry and set out its plan for physical AI. It is one pillar of the three mega-projects the government calls Korea's great leap, alongside AI data centers and next-generation semiconductors. At the heart of it is a Physical AI Training Center to be built from 2027, meant to produce and put to use the high-quality data that will teach robots and AI models. The outline known earlier from the budget stage describes an environment in which robots and sensors record real movement, and companies and research institutes turn those recordings into behavior data to train and verify models. The National IT Industry Promotion Agency handles construction. Field demonstrations run alongside it in real workplaces such as logistics, in partnership with Korea Post.

The money appears as new line items in the 2027 government budget bill: ₩40 billion to build the training center, ₩55 billion for core technology development and ₩26 billion for demonstration and rollout tied to Korea Post. Across the three mega-projects the total rises from ₩298.1 billion in 2026 to ₩795.6 billion in 2027, and the ministry's overall AI budget is set at ₩9.4 trillion. None of this is settled money yet. The 2027 budget goes through National Assembly review before the numbers harden. What the government has drawn is a domestic full-stack ecosystem running from semiconductors through AI models and data to robots. We went through how those three line items split in an earlier article.

Two places for making data opened in the same month, then, and what they will hold barely overlaps. One is a workplace where mail moves through. The other is a house where people fall asleep. In the workplace the same task repeats along a fixed path, and at home the same task happens once, in a different shape in every household. The bodies running them differ, so do the robots that go in first, and so does the channel the money flows through.

The Seoul city and SNU agreement announcement and the ministry briefing are each set down as given and placed side by side. Pairing the two programs appears in neither announcement, and the framing is this article's own.

That the two kinds of data do not overlap is no bad thing in itself. Records from a living room and records from a logistics warehouse are both needed, and neither stands in for the other. What neither announcement says is how the records coming out of the two sites are to be made to mesh, or who sets a common format.

Three parties make records at once inside that one living lab: the demand-side institution that raised the problem, the company that brought the robot in, and the laboratory doing the verification. The video, the trajectories and the failure cases that pile up while the sensors run belong to which of the three? The agreement announcement has no answer. Space, schedule and the scale of support are written down to the number, and the ownership structure of the records is blank.

The question of format slips past even more quietly. Five companies a year makes twenty-five over five years, and if each leaves with its records in its own format, what remains is twenty-five pieces that cannot read one another. Even for records shot in the same living room by the same sensors, a different labeling scheme, a different coordinate system or a different definition of failure puts them out of reach of the next company. Sharing a space and sharing data are two different things, and the first does not produce the second on its own.

Three things in the next announcement will show whether this living lab is heading toward a shared asset.

• Whether the contract writes down who holds the records and the right to use them, and how that is divided. The range a company takes away, the range the laboratory keeps and the range of consent given by the person being cared for have to be written separately for there to be anything to go back to when a dispute arises
• Whether the format for keeping records is settled first. If whatever the first consortium uses becomes the standard in practice, then picking one trial operator is the same act as picking the format
• Whether only the successful attempts are kept, or the failed ones too. The moment a robot catches on a threshold and drops the cup is the most expensive material there is for teaching the next robot, and verification reports usually record nothing beyond pass or fail

Asking who owns the records a robot makes is not new here. Pebblous has written about the rights attaching to robot data under Korea's Industrial Digital Transformation Act and the EU Data Act, and about a robot dataset released with its failures annotated. The same questions attach, unchanged, to the records that will pile up in that living room in Yangjae.

Thank you for reading this far. The date of the agreement and who attended, the distinction between the 612-square-meter center and the 338-square-meter first-phase living lab, the schedule of construction in the first half of 2027 and verification in the second, support for five companies a year, the five kinds of robot to be tested and the three axes of performance, safety and user acceptance, the division of roles between Gwanak and Yangjae, H-PAIR's ₩95 billion in state funding and the scale of its participation, the plan to open MORPHI and MOSAIC, and the structure of the Physical AX Consortium were confirmed in the agreement announcement from Seoul and SNU's College of Engineering and in the reporting that carried it. The plan for the ministry's Physical AI Training Center and the Korea Post partnership follow the October 29 briefing, and the 2027 budget items and amounts come from reporting on a government budget bill still facing National Assembly review. Section 3, which draws a data problem out of the choice to put everyday support first, the table in Section 4 setting the two programs side by side, and Section 5's questions about who owns the data and in what format it is kept, are this article's own. If you keep records from a demonstration site, we would be glad to hear what format you keep them in.

Pebblous Data Communication Team
November 2, 2026

Primary Sources — Agreement and Government Announcements

• 1.Bang Eun-ju. (2026). "서울대-서울시, 피지컬AI 협력...내년 양재에 센터 개설" [SNU and Seoul Cooperate on Physical AI, Center to Open in Yangjae Next Year]. ZDNet Korea, 2026-11-01.
• 2.Woo Tae-seon. (2026). "서울대 공대·서울시, 양재에 피지컬 AI 실증 거점… 'H-PAIR 피지컬 AX 센터' 세운다" [SNU Engineering and Seoul Build a Physical AI Demonstration Base in Yangjae]. Korea Data Economy News, 2026-11-02.
• 3.Oh Byung-hoon. (2026). "정부, 피지컬AI 실증·데이터 사업 본격화…내년 트레이닝센터 구축" [Government Moves Ahead on Physical AI Demonstration and Data Projects, Training Center Next Year]. DDaily, 2026-10-29.

Prior Pebblous Coverage

• 4.Pebblous Data Communication Team. (2026-10-14). "Korea Builds a Physical AI Training Center, but Nobody Has Set the Data Format."