[AI Pick] The era of AI moving factories is coming... Speed of physical AI commercialization

Development of manufacturing-specific AI model and factory operating system… Industry-academia-research commercialization strategy discussion

Promoting safety standardization and digital twin verification… One Predict’s commercialization goal next year

(Seoul = Yonhap News) Reporter Oh Ji-eun = Ahead of the era of physical AI, in which artificial intelligence (AI) autonomously controls equipment and robots at manufacturing sites, domestic industry, academia and research institutes are accelerating the development and commercialization of factory automation technology.

From building a manufacturing-specific AI model and factory operating system (OS) to safety standardization and prior verification using digital twins, the technical tasks and development directions for the introduction of physical AI into actual industrial sites are being specified.

◇ Promoting factory autonomy with manufacturing foundation model and agentic AI

At the 'Physical AI Tech Week 2026' technology insight session held at COEX in Gangnam-gu, Seoul, hosted by the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) on the 14th, the next-generation manufacturing ecosystem strategy that links factory operation, standards, and virtual verification (digital twin) into one was discussed.

In the session chaired by Professor Kim Soon-tae of Chonbuk National University, experts from academia, industry, and standardization organizations gathered to present a detailed roadmap ranging from the manufacturing foundation model to the factory operating system (OS) and international standardization.

Byung-dong Yoon, CEO of One Predict, presented AI Ready Data, Manufacturing Foundation Model, Agentic AI, and Physical AI as the four major axes for factory autonomy.

CEO Yoon said, "There is an overabundance of data in the field, but there is a severe lack of 'useable data' that is contextualized and structured based on facility and time so that AI can learn and infer. We need to move away from the existing inefficiency of creating individual models for each facility or task and move toward building a common foundation model and then fine-tuning it with a small amount of data."

In particular, high-frequency time series data, in which thousands of signals are generated per second, was cited as Korea's strength.

CEO Yoon explained, "Although global big tech has already dominated text and images, Korea is the world's best in the high-frequency time series field. We confirmed high accuracy as a result of applying our foundation model, which was learned from 70 million pieces of data, to semiconductor logistics robots and food and beverage (F&B) processes."

He added, “We plan to undergo factory field verification by the end of this year and introduce a commercial model early next year.”

◇ Physical AI safety standardization begins in earnest… Pre-verification with digital twin

The establishment of a standardization and safety standard system for the introduction and commercialization of physical AI in actual factories is also in full swing.

Kim Jae-geun, director of the Korea Telecommunications Technology Association (TTA), emphasized the need for industrial standards, saying, “The LLM on the screen can be relearned if an incorrect answer is given, but errors in physical AI directly lead to equipment damage or safety accidents.”

TTA is promoting the standardization of the Physical AI Asset Management Shell (PAAS) to be installed in the factory OS in connection with the Jeonbuk Physical AI R&D project.

It is a method that combines the existing International Electrotechnical Commission (IEC) equipment information standard with a sub-model dedicated to physical AI, such as autonomy level, action tolerance, and collaboration protocol.

Professor Seo Gyeong-min of Hanyang University, who dealt with robot-based physical execution, highlighted the value of digital twins in sites where various types of robots and workers coexist.

Professor Seo said, “The cost of failure can be significantly reduced by prior learning and verification in a virtual environment before actually deploying robots in manufacturing and logistics sites.”

He went on to suggest, “A sophisticated simulation architecture that overcomes the Real-to-Sim gap that occurs when transferring real data to a virtual environment, the Sim-to-Sim gap that occurs between learning and verification environments, and the Sim-to-Real gap that occurs when a model verified in the virtual environment is transplanted to a real robot must be supported.”