[Edaily Reporter Han Gwang-beom] The government has set out to build an independent physical AI ecosystem, spearheaded by the “World Model,” which learns real-world physical laws in a virtual environment. The strategy aims to secure a leading position in the global AI hegemony race by localizing simulator technologies—which had previously relied heavily on foreign imports—based on large volumes of synthetic data.
On the 9th, the Ministry of Science and ICT held a kickoff meeting for the “Physical AI Leading Technology Development Project” at LG Science Park in Magok, Seoul, officially launching a large-scale national project to secure core foundational technologies. This project is one of the key initiatives under the government’s “K-Moonshot” national AI innovation project announced earlier this year, with a total of 34 billion won to be invested over the next two years.
The event was attended by representatives from major companies—including #LGElectronics, #KT, #MaumAI, #Robotis, Holiday Robotics, Crowdworks, and Alchela—as well as researchers from KAIST and Seoul National University, and representatives from industry, academia, and research institutions such as the Telecommunications Technology Association (TTA), who discussed plans for developing Physical AI technology and fostering an ecosystem.
Kim Young-joon, Head of the LG Electronics AI Research Lab and the consortium’s research director, emphasized the need to secure manufacturing-based physical AI technology and presented “100% localization” as a core goal.
“Current robots lack the ability to perform complex tasks continuously and face the limitation of having to relearn data every time a new task is assigned,” said Director Kim. “To solve this, a world model that internalizes an understanding of the physical world is essential.”
He cited “manufacturing site data” as a competitive advantage that sets Korea apart from global Big Tech companies. “Factory data and equipment operation experience are assets that cannot be easily obtained from external sources,” he emphasized, adding, “We will leverage the strengths of Korea’s manufacturing industry to develop a competitive, proprietary model.”
He added, “We will create a virtuous cycle where we continuously refine the world model through trial and error gained from applying it to actual industrial sites, rather than remaining confined to the laboratory.”
“Full Localization of Models, Hardware, and Simulators”The ultimate goal of this project is to establish a Physical AI pipeline based on proprietary technology.
LG Electronics is developing the world model based on its own super-large AI model, ‘EXAONE,’ while KT is leveraging its generative AI ‘Mi:dm’ and platform operation experience to connect it to a robot behavior model (RFM).
Robotis is responsible for robot hardware and control technology, while Holiday Robotics is tasked with developing a domestically produced multi-physics simulator engine that precisely implements the laws of physics. Research teams from KAIST and Seoul National University plan to secure technological competitiveness by concurrently developing world models using different methods, such as 3D reconstruction, video generation, and JEPA.
Director Kim explained, “This project is a bridge project aimed at rapidly implementing a minimum viable product (MVP) and conducting repeated field validation,” adding, “Our goal is to localize the entire technology stack and provide the ‘K-World Model Implementation Recipe’ to the domestic industry.”
“Tobe released as
open source so that SMEs can also utilize it”The development results are planned to be released as
open sourceto expand the domestic AI ecosystem.
Kim Young-jun, Head of LG Electronics’ AI Research Lab, stated, “We will release the R&D results via GitHub to support SMEs without large-scale data or infrastructure in achieving automation innovation using only small amounts of field data,” adding, “We will also provide tools and interfaces to enhance usability.”
He also raised the need for the construction of dedicated infrastructure at the national level.
He emphasized, “Physical AI must process far more video and simulation data than text-centric large language models (LLMs), requiring massive storage space and computational resources.” He added, “If a national-level integrated infrastructure for storage, training, and validation—combining ultra-high-speed storage, GPUs, and validation environments—is established, we will be able to produce much more competitive research results.”
