[E-Daily Reporter Shin Yeong-bin ] LGELECTRONICS(066570)is training artificial intelligence (AI) robots to perform the task of transporting steel plates—previously done by humans—at its washing machine factory in Tennessee, U.S. While developing technology that enables robots to handle heavy workpieces in actual manufacturing processes, the company is also increasing the amount of data required for training to improve operational accuracy.

Baek Seung-min, head of LGELECTRONICS’ Advanced Robotics Research Lab, presented the current status of robot learning technology development on the 15th at the “2nd Korean Robotics Society Expert Symposium” held at the Korea Institute of Science and Technology (KIST) in Seongbuk-gu, Seoul.

The case study unveiled by LGELECTRONICS involves the steel top panel handling process at a washing machine factory in Tennessee, U.S. When a worker hangs a washing machine top panel—weighing approximately 4 kg—on a coating hanger, it is fed into the equipment where the coating process takes place. The process of removing the coated steel panel and placing it back down is also currently performed by a human operator.

LGELECTRONICS is developing technology to have robots perform this task. However, it was difficult to implement the task using the remote-control method typically employed for training robots.

The robot must maintain the appropriate force when gripping and lifting the steel plate with both hands, but it was difficult to control this with precision using remote-control equipment. The operation speed was also too slow for application in an actual production process.

Director Baek explained, “The data collected via remote control is too slow,” adding, “To grip a washing machine steel plate from both sides, the robot needs to sense the force in both hands and provide a certain level of support, but remote control equipment is not well-suited for that.”

To solve this problem, LGELECTRONICS first utilized existing robot control technology. This method involves using a camera to determine the position and orientation of the steel plate, planning the motion, and then having the robot perform the task. After implementing these automated motions, the company used the resulting data to train the VLA model.

Director Baek unveiled a video showing a robot lifting a steel plate and hanging it on a hanger on-site. “Handling actual workpieces that are heavy and subject to various constraints is extremely difficult in manufacturing processes,” he said, noting that the technology is being developed with real-world process applications in mind.

LGELECTRONICS is also collaborating with NVIDIA to address the shortage of data required for robot training. This approach involves using NVIDIA’s world foundation model “Cosmos” to generate synthetic data for various environments based on real-world task data.

Director Baek explained that it is difficult to reproduce the forces or sensory information generated when a robot comes into contact with an object using synthetic data alone. Therefore, the company is training robots by supplementing real data with synthetic data that simulates changes in the surrounding environment.

“We found that simply mixing synthetic data augmented with real data improved task execution accuracy,” he said, adding, “I will discuss this in more detail later as part of our collaboration with NVIDIA.”

LGELECTRONICS is also collecting robot operation data at its Yangjae Data Factory in Seoul, where it has recreated manufacturing, commercial, and residential environments. The company is developing technology to train robots by replicating actual factory work environments and then applying this training to production sites.

The remaining challenge is reducing training time. According to Director Baek, it took several months to teach robots to handle steel plates at the Tennessee plant. He explained that if each new process requires a long training period, it will be difficult to ensure the economic viability of introducing robots.

LGELECTRONICS plans to advance technologies such as few-shot learning—where robots learn new tasks with minimal demonstrations—zero-shot learning, which enables robots to perform tasks without additional training, and the ability to apply existing task capabilities across different factory environments.

Director Baek emphasized, “A robot must be able to perform the tasks it learned today even when the environment changes tomorrow,” adding, “Ensuring that tasks learned by one robot can be performed by other robots is a challenge that the Robot Foundation Model (RFM) must address.”