LG has four rooms where robots train. Every hour flows to Nvidia.

In a replicated Seoul apartment, LG's CLOiD robots practice folding and wiping. In a simulated Tennessee washing-machine plant, the same units stack and assemble parts. In a logistics yard, they pick. In a hand-training room, they grip. Every hour of motion in all four spaces flows into Nvidia's stack, where it becomes training data for the next generation of the same robots.

LG and Nvidia signed the partnership in Santa Clara on August 13. The model being trained is what LG calls a Robot Foundation Model, or RFM, the company's "physical AI" model for robots that act in the real world, distinct from chatbots or image-only systems. The RFM is being trained on motion data that does not yet exist in useful quantity, generated by robots that are not yet good enough to generate it.

The four rooms sit in three different companies. The first, the Seoul apartment, sits inside LG Electronics. CLOiD is LG's home and service robot line. The second, the simulated Tennessee plant, is also an LG Electronics project, with the digital twin built to match the layout of the company's actual washing-machine factory. The third, the logistics yard, is run by LG CNS, the group's IT-services affiliate. The fourth, the robotic-hand training space, sits inside LG Innotek, the group's components affiliate.

Each room produces a different kind of motion data. None of it is useful on its own: a robot that can fold a towel cannot yet stack a transmission housing, and a gripper tuned for a logistics tote will fail on a glass panel. The point of running four rooms at once is volume and variety, so that when the data is fed back into a single model, the model can learn things no single room could teach it.

That is the "data flywheel" LG keeps using in its corporate communications. The term is borrowed from the way large language model labs describe their own training pipelines: more usage produces more data, more data produces better models, better models produce more usage. LG's version is physical. The robots in the rooms get better. The better robots generate cleaner data. The cleaner data trains a better RFM. The better RFM is what shows up in a future CLOiD, or in a future LG CNS logistics system, or in a future Innotek component.

Nvidia's role inside that loop is the stack, not the model. The named components are Omniverse for the simulated Tennessee plant and other digital twins, Cosmos for the open world foundation models that give robots a baseline sense of how objects behave, and Isaac for the robotics development platform that handles training and deployment. LG runs the rooms and the data aggregation. Nvidia provides the substrate.

The Data Factory at Yangjae, in Seoul, is where the loop is supposed to close. LG says the facility is being built to produce 100,000 hours of training data by year-end. That number is an LG-set internal target, not an independently measured outcome. The Data Factory itself is described as "under construction," and no ship date for a consumer-facing CLOiD model has been published.

Lyu Jae-cheol, the CEO of LG Electronics, framed the deal as "One LG" synergy, the parent group coordinating Electronics, CNS, and Innotek so that the same training loop serves a consumer robot, a factory robot, and a components business. The same release pitches LG as a future "comprehensive robotics solutions provider," a phrase that reads as ambition rather than current position.

The announcement does not settle the harder questions. It does not name a product, a price, a customer, or a launch window for any of the robots the data is meant to train. It does not show a benchmark. It does not say what happens to the loop if Nvidia's pricing or access terms change.

It does name the rooms. By year-end, the 100,000 hours either exists in Yangjae or it does not.