LG and Nvidia put a number on physical AI: 100,000 hours of robot data in one year

LG Electronics said chief executive Ryu Jae-cheol and Madison Huang, Nvidia's senior director of omniverse and robotics, met on the 18th at LG's R&D campus in the Yangjae area of Seoul and agreed to train systems on 100,000 hours of robot data, described as roughly 12 years' worth, within the year ^[1][3][4]. That lands four days after chairman Koo Kwang-mo signed a memorandum of understanding with Nvidia's chief executive in Santa Clara to develop a next-generation bipedal humanoid, and it is the more informative of the two events, because a data target implies equipment, floor space and a deadline ^[2][14].

The venue carries the plan. LG's DataFactory, being built at the same campus as what the company calls a robot bootcamp, spans four floors from one basement level to the third, with 10,000 square meters of floor area, and is to receive several hundred robots including LG CLOi within the year to gather data and verify performance ^[4][9][10]. Nvidia's side of the arrangement is software and platforms: the Omniverse Library, the Cosmos open-world model and Isaac open robotics, applied on site ^[11]. LG says the output feeds its Robot Foundation Model, the robot brain ^[12].

The arithmetic deflates the headline framing and improves the operational one. 100,000 hours is about 11.4 years of continuous around-the-clock running, so "12 years' worth" is elapsed clock time rather than working shifts ^[1]. Spread across 300 robots, the midpoint of "several hundred," each machine needs roughly 333 hours, about 42 eight-hour days of logging ^[2]. That is not a scheduling feat. It is a fleet and a building: at 300 units, the 10,000 square meter facility allots about 33 square meters per robot ^[3]. The competitive variable being purchased here is throughput of physical trials, not model cleverness.

LG's stated reasoning is that robots, unlike text or image models, need unstructured data such as the movements and know-how of on-site workers, which makes data volume a decisive battleground ^[8]. An LG official said combining the high-quality data the company has accumulated over decades at manufacturing and logistics sites with Nvidia's robotics solutions "is expected to be the secret to building a world-class data cycle that sets it apart from other companies" ^[13]. Ryu framed the effort as securing competitiveness through "One LG" synergy and emerging as a total solution provider in robotics ^[6]. The supporting scaffolding is already in place: affiliate heads including LG CNS chief Hyun Shin-kyun and LG Sciencepark head Chung Soo-heon attended the meeting ^[7], LG set up a Robotics Business Center last month reporting directly to the CEO ^[16], and it has begun mass production of AXium, an actuator serving as a robot joint ^[17].

Three dated checkpoints are worth tracking. The humanoid from the Koo-Huang MOU is targeted for unveiling in the first quarter of next year ^[14]; CLOi is to be field-tested at LG's Tennessee plant within the year, with an AI factory to follow next year ^[15]. The open question is disclosure: whether LG reports hours actually logged, and how it splits real-world capture from what Cosmos and Omniverse generate in simulation, since the announced target as described does not distinguish the two ^[4][11]. Madison Huang called the outcome of the roughly two-hour meeting "incredible" ^[5], which is a reaction, not a delivery record.