LG’s Seoul Lab Is Minting 100,000 Hours of Robot Data

LG Electronics is packing a 10,000 square meter Seoul lab with hundreds of robots to mint 100,000 hours of training data by the end of 2026. Senior Nvidia officials walked that still-unfinished Yangjae floor on August 18. LG Corp chairman Kwang Mo Koo had signed the robotics and AI factory deal at Nvidia headquarters in Santa Clara on August 13.

The bipedal humanoid is still dated to the first quarter of 2027. The first machine that has to earn its keep is a wheel-based CLOiD on a Tennessee washer line, scheduled in 2026.

Seoul’s Data Factory Is Racing to 100,000 Hours

LG hosted the Nvidia visit at a robotics lab under construction on the Yangjae R&D Campus, and it published the specs the same day. The Yangjae Data Factory in Seoul covers 10,000 square meters across four floors, one basement and three above ground, and it is supposed to be fully running by the end of 2026.

By then LG says the building will hold several hundred robots. The 100,000-hour target mixes hours collected on those machines with hours minted in software through Nvidia Cosmos world models. LG calls that combined pile roughly 12 years of data, and it wants the lot fed into its own Robot Foundation Model, the in-house brain meant to sit under later humanoids.

THE SEOUL LAB ON PAPER

• Floor plate: 10,000 square meters on four levels at the Yangjae R&D Campus.
• Robot count: several hundred units planned inside the building by the end of 2026.
• Data target: 100,000 hours of real plus synthetic training data by the same deadline.
• Command line: a Robotics Business Center set up in July 2026 that reports to the LG Electronics CEO.

Lyu Jae-cheol, the LG Electronics chief executive, said the group would use One LG hardware plus outside partners to become a full robotics supplier, not a one-off demo shop. That is a different job from showing a walking machine on a stage in early 2027, and the calendar treats it as the job that starts now.

A Replica Home and a Replica Tennessee Line Share One Building

The lab is not a single open hall. LG has already stood up task rooms so CLOiD units can repeat the same work, log it, and send the traces into Nvidia’s robotics stack for extra synthetic copies. Nvidia Omniverse libraries, Cosmos, and the open Isaac tools are the path from those rooms to a trained policy.

FOUR ROOMS INSIDE THE YANGJAE LAB

• Home replica: CLOiD units repeat cleaning work in a mocked-up house so household motion is on tape before a customer hallway is.
• Tennessee replica: a mock washer plant, modeled on LG Electronics’ U.S. line, where the same robots move, stack, and assemble parts.
• Logistics bay: robots run through LG CNS automation layouts rather than a show-floor obstacle course.
• Hand cell: LG Innotek trains robotic hands in a dedicated space, feeding finger and grip data into the same pile.

LG’s pitch is a data flywheel: decades of plant and logistics traces, plus these new robot hours, get cleaned, copied, and trained until the next robot shift is cheaper than the last. That only works if the rooms stay full and the synthetic copies stay honest, which is why the 100,000-hour clock is the number the company put on 2026 rather than a joint count of humanoids sold.

Why the Washer Plant Matters More Than the Humanoid

The August 13 statement still leads with a next-generation bipedal machine. It will use Nvidia Jetson Thor for onboard compute, the Isaac GR00T humanoid foundation model for reasoning, and Nvidia Halos for Robotics as the safety stack. Public unveiling is set for the first quarter of 2027.

Before that date, LG says it will put wheel-based CLOiD robots on the washing machine line at its Tennessee plant in 2026. If the trial holds, the company says it will widen the same design to other plants, homes, and commercial sites. The Seoul lab already contains a copy of that washer floor, so the first “factory” data is being staged in Korea, then graded in the United States.

Building on years of collaboration, LG and NVIDIA are combining LG’s leadership in product engineering and manufacturing with NVIDIA technology to accelerate the next era of robots, AI factories and autonomous vehicles.

Jensen Huang, founder and CEO, Nvidia, in the August 13 statement

A stage walk in 2027 will not tell anyone whether a wheeled unit can last a shift beside a washer drum. The Tennessee line is the first place that question gets a shop-floor answer, which is why the 2026 trial is the harder date on the sheet.

Congratulations to @LGE_Global on its new Data Factory in Seoul. 🎉

NVIDIA and LG are expanding their robotics collaboration to accelerate physical AI development through high-quality data for robotics and industrial AI. pic.twitter.com/hlW1MguOd7

— NVIDIA Robotics (@NVIDIARobotics) August 18, 2026

One LG Builds the Body, Nvidia Sells the Brain

LG is not buying a finished humanoid from Santa Clara. The group is trying to keep the metal in-house and rent the mind. LG Electronics is down for actuators, LG Innotek for sensors, and LG Energy Solution for batteries. LG CNS is building PhysicalWorks as the on-site data platform for collection, synthetic generation, training, and checks.

Nvidia’s side of that split is consistent across the June and August papers: Jetson Thor on the robot, Isaac GR00T as the open humanoid model, Halos as the safety layer, and Cosmos plus Isaac as the data mill. The two firms also said they will run a joint task force across research, site trials, and sales, which is the closest thing in the file to a staffing plan.

WHO PUTS WHAT ON THE ROBOT

That split is also the limit. If GR00T and Thor become the default brain for every serious humanoid, LG’s metal has to stay cheaper, tougher, or easier to service than the next body on the same chips. The closed loop of LG plant data is the only piece a rival cannot order from the same catalog.

Hyundai’s Atlas Is Already Walking in Public

LG is not the only Korean group trying to put a human-shaped machine on a factory floor, and it is not first in public. Hyundai Motor Group has said it plans to deploy more than 25,000 Atlas robots from Boston Dynamics across Hyundai and Kia plants, and to build as many as 30,000 humanoids a year by 2028, with an early run at the Georgia Metaplant in 2028.

Atlas already walks in front of cameras. LG’s biped does not, and the 2027 unveil is still a target, not a shift roster. Nvidia can live with both outcomes. The same company is selling simulation, onboard compute, and foundation models into more than one Korean industrial plan, so it gets paid whether the winning body is bolted together in Changwon or in Georgia.

The sharper problem for LG is time. A washer-line CLOiD in 2026 can still be a useful industrial robot if the biped slips. It cannot rewrite the fact that a rival already has a walking platform and a unit target with a year on it. Factory data is the counter, and it only counts if the Tennessee trial produces traces that a foundation model can actually use.

Cheonan Gets an 80 MW AI Factory in 2028

The other product in the August papers is not a robot. It is a computer plant. LG said it will put up an AI factory reference site on Nvidia Vera Rubin hardware in the first half of 2027, then an 80 MW LG AI Factory in Cheonan, in Chungcheongnam-do, in the first half of 2028, to train physical AI and robot foundation models.

That track started in public on June 8, when Koo and Huang held a top management meeting at the LG Twin Towers in Yeouido. Nvidia’s June write-up described an AI factory platform for LG Group that ties model work, physical AI data, robot simulation, edge rollout, and factory-scale digital twins into one workflow. The August deal put dates and a megawatt figure on that sketch, after the physical AI stack they locked in during the summer meetings.

LG Electronics is also in the cooling business Nvidia needs for those halls, including coolant distribution units and cold plates, plus prefabricated modular halls aligned with the Nvidia DSX layout. LG Energy Solution is working on 800 volt direct-current power kit. LG Uplus and LG CNS are on the hook for large DSX-style sites. The group, which says it operates in more than 60 countries and books USD 140 billion in annual revenue, wants to sell that bundle as a turnkey package to other big computing customers, not only to run it for its own robots.

THE BUILD DATES LG PUT ON PAPER

1. June 8, 2026: Koo and Huang meet in Yeouido and widen the work across robots, AI factories, and mobility.
2. August 13, 2026: They sign the memorandum in Santa Clara and name the biped, the Tennessee trial, Vera Rubin, and Cheonan.
3. August 18, 2026: Nvidia officials tour the Yangjae lab still under construction.
4. End of 2026: Yangjae is supposed to run; several hundred robots and 100,000 hours are the stated marks; CLOiD is due on the Tennessee washer line.
5. First quarter of 2027: public unveil of the bipedal humanoid on Jetson Thor, Isaac GR00T, and Halos.
6. First half of 2027: Vera Rubin AI factory reference site.
7. First half of 2028: 80 MW LG AI Factory in Cheonan.

Huang said in June that Korea’s mix of manufacturing, mechatronics, and AI could make robotics a major growth sector for the country. Koo said on August 13 that the tasks in AI factories and physical AI had become clear, and that LG would push adoption by building reference sites. Those are still statements of intent. The 80 MW hall is a 2028 building. The robot that has to work first is a wheeled unit beside a washer in 2026.

The Dates Only Count If the Robots Work

None of the August paper is a shipping manifest. It is a memorandum, a joint task force, a lab under construction, and a set of dates. The 100,000 hours include synthetic copies. The humanoid is an unveil, not a production forecast. Cheonan is a first-half 2028 target on modular construction, not a live cluster.

What the file does fix is order. Data collection and a washer-line trial are dated to 2026. The walking biped is dated to early 2027. The big training plant is dated to 2028. Nvidia can supply the same brain to whoever else can pay for Thor and GR00T. LG’s remaining edge is the closed loop of its own plants, starting with a replica of Tennessee inside a Seoul lab and then the real line in the United States.

If CLOiD lasts a shift on that washer floor, the 100,000-hour pile starts to look like an industrial asset. If it does not, the first-quarter 2027 photo will still be a photo.