LG spent about a year teaching a humanoid one washer-line task in Tennessee; the demo took days

LG Electronics has spent about a year testing a humanoid robot on a single job at its washing-machine plant in Tennessee: lifting the machine's top plate and hanging it on a coating hanger. Park Jeong-seop, who leads robotics intelligence at LG, presented the trial on Nov. 9 at NVIDIA AI Day Seoul 2026 at COEX. The job was picked because workers who lift the plates over and over build up fatigue and musculoskeletal strain, iNews24 reported.

The most useful number came in the Q&A. A short demo was possible within days, Park said, according to Chonji Ilbo; the hard part was keeping the robot working as part positions drifted and the lighting changed from run to run. That gap is the cost of a factory humanoid. The demo is cheap; the rest of the year went into making it repeat.

Each failure goes back into simulation

LG split the work into five stages. Engineers analysed skilled workers' motions, then collected teleoperation and simulation data and trained NVIDIA's Isaac GR00T model on it. They used NVIDIA's Cosmos video model to generate synthetic backgrounds so the robot would cope with light and surroundings, and built a near-copy of the line in Isaac Sim to shrink the sim-to-real gap. TensorRT cut the time the robot takes to decide its next move. When the robot fails on the floor, the team recreates the failure in simulation and retrains, iNews24 reported.

"On a real factory floor there are a great many variables every day, depending on the production plan," Park said. "Getting performance that holds up through those changes is the important homework." Lee Jong-hwan, an NVIDIA solution architect for physical AI, said in the same session that a robot trained only on successes ends up replaying memorised paths, so it also needs data on recovering after a failure.

The next 500,000 hours are for repeats

LG plans to reach 500,000 cumulative hours of training data by the end of the year through its LG Data Factory, and to verify the robot again and again in the plant and in simulation. The goal, Newstop Korea reported, is better repeat accuracy, adaptation to changing conditions and recovery after errors.

The strongest case against reading too much into the year is that much of it may have gone into building the pipeline, and a pipeline is reusable. LG put its wheeled CLOiD robot into the Tennessee plant in August and plans to unveil a bipedal humanoid built on Isaac GR00T in the first quarter of next year, iNews24 reported. If the second task on the line takes weeks rather than a year, the year was an investment; if it takes another year, it was the price per task.