“LG’s physical artificial intelligence (AI) aims to make an ‘impact’ in the field by 2027.”

This was the point emphasized by Kim Seung-hwan, Executive Vice President and head of the Physical Intelligence Lab at LG AI Research Institute, in an interview with the Dong-A Ilbo on the 9th at the institute in Gangseo District, Seoul. LG AI Research Institute is conducting research so that the physical AI currently under development can be deployed at manufacturing sites of group affiliates next year to generate results such as production cost reductions and yield improvements.

● LG Group accelerates physical AI

In January this year, LG reorganized the “Vision Lab” of LG AI Research Institute, which oversees the Group’s AI research, into the “Physical Intelligence Lab.” The organization was expanded by adding personnel in charge of reinforcement learning (a learning method in which AI accumulates experience through trial and error) from the Data Intelligence Lab to the existing Vision Lab staff. The unit that had focused on image and video recognition AI has been restructured into a dedicated organization for developing “behavioral models” that research robots and smart factories. Kim stated, “The starting point of robot technology is giving robots eyes,” adding, “The task is to expand images into space, then further into space-time, and ultimately develop a ‘robot foundation model.’”

The model under development by the Physical Intelligence Lab at LG AI Research Institute is different from large language models (LLMs), such as “Exaone,” through which the institute has demonstrated results. While existing models focus on understanding the surroundings through processes such as seeing, hearing, and reading, behavioral models infer how to grasp objects or avoid obstacles based on that perception. They also determine how joints should move according to the situation. Kim said, “‘Sim2Real,’ which implements in the real world the behaviors learned in virtual reality, is the core of behavioral model development,” emphasizing that “the key to success or failure will be how closely the constructed data resemble reality.”

● Manufacturing as the key to Korea’s physical AI

LG is looking to manufacturing as the solution for building physical AI behavioral models. Kim explained, “Through decades of manufacturing experience, LG has accumulated on-site data that others do not have, and is utilizing them,” adding, “LG Innotek’s case of automating the product defect inspection process is representative.”

LG Innotek introduced “AI vision inspection” into its semiconductor substrate production process, building a system in which AI identifies defects. This enabled the company to reduce both delivery time and inspection personnel by about 90%.

There is a forecast that manufacturing sites where physical AI is applied will advance beyond simple “automation” to “autonomization.” This is because factories will become places where physical AI synthesizes visual information received through vision sensors and tactile information perceived by fingertip pressure sensors to derive the most appropriate action in the moment and act proactively. This is a different concept from automation, in which robots merely repeat predetermined movements based on pre-entered information.

Kim believes that Korean companies have greater potential for development in the physical AI market than in the LLM market, where investment has been somewhat slower compared with US and Chinese AI companies. He stressed, “Robot foundation models being released by global big tech companies are still at an early stage,” and “If Korea’s on-site manufacturing experience and infrastructure are combined with foundation model development capabilities that are not inferior to those of global companies, there is ample chance of success in the physical AI market.”

Popular News