Korea was never a country that only does software (SW). It is a country strong in things that move, such as factories, equipment, automobiles, shipbuilding, and infrastructure. Physical AI (AI with a physical form, such as Autonomous Driving cars or robots) is a game changer that can leverage this very strength."
In a recent interview, Jang Young-jae, head of the manufacturing Physical AI Research Institute at Korea Advanced Institute of Science and Technology (KAIST), told corporations and government official seeking manufacturing innovation in the Physical AI era, "We must break away from the defeatism that, because we fell behind in the large language model (LLM) race, we should give up on artificial intelligence (AI)," adding, "Physical AI is not a theory on a desk, but a technology that matures through trial and error in the field." Considered the top domestic expert in Physical AI, Jang serves as the overall research lead for the "Physical AI core technology demonstration (Jeonbuk specialized model development)" project, a national initiative worth around 1 trillion won. Jang said, "In that it attempts to build an industrial ecosystem model for Physical AI, this project will be an important turning point for the development of Physical AI in Korea." The following is a Q&A.
What is a world model?
A world model (an AI model that learns physical laws to simulate reality) is a neural network model that compresses and represents within itself the dynamics of the real world—namely spatial structure, physical laws, and changes over time. It is the brain of a factory that can predict what will happen next when a certain action is taken in the factory, and the concept is that the simulation engine is contained in a single model.
In manufacturing settings, world models play two roles. First, in a Digital Twin (a digital model identical to a real building or factory), they quickly learn equipment and processes through hundreds of thousands of virtual experiments, giving Physical AI "real-world instincts" before it is deployed in the factory. Second, they continuously incorporate real-time data from the field to infer the current and next states, creating a foundation that allows autonomous responses after detecting process condition changes or equipment anomalies in advance. Ultimately, world models will turn the entire factory into one massive Reinforcement Learning environment and become the core engine for Physical AI to evolve into a "self-learning autonomous manufacturing system.""
What role will Physical AI play in manufacturing innovation?
Autonomous manufacturing led by Physical AI means a system that perceives the physical world of the factory, makes its own judgments to choose optimal actions, and even learns. Based on real-time data from sensors, cameras, robots, and equipment, it understands process states and replans, aiming for a factory that autonomously adjusts production conditions and work sequences without human intervention. The core of Physical AI manufacturing is the evolution from automation, which repeats predefined tasks, to autonomy, which understands situations and changes work on its own."
Specifically, what will it change manufacturing into?
Until now, manufacturing has mainly been perceived as "an industry that makes and sells products." As Physical AI matures, operational know-how—namely how to operate efficiently, safely, and flexibly—will become a key competitive edge, more so than product making itself. A new manufacturing model may also emerge that exports such know-how in the form of software and platforms. For example, if a certain country or corporations build a Physical AI platform for operating battery factories, automobile factories, and logistics centers, they can export "operational technology" by replicating this platform to factories and cities in other countries."
What industries will Physical AI be applied to?
I expect three broad categories. First, high-tech manufacturing fields with complex processes and heavy capital expenditure, such as semiconductors, batteries, and precision parts. In these domains, the expense from downtime (unexpected machine stoppages) and quality variation is massive, so the economic effect of factory-level optimization and autonomous operations is large. Second, the automobile and mobility, logistics, and warehousing industries that are directly tied to global supply chains. This field can quickly secure a competitive edge by integrating Physical AI component technologies such as robots, Autonomous Driving, and predictive maintenance. Third, the medical, care, and service robot field, where aging and population decline have major spillover effects. There is an inevitable demand to directly replace and supplement human labor."
The AI utilization rate in domestic manufacturing is 10%. What tasks are needed to raise the success rate of Physical AI?
If only about 10% of manufacturing AI has succeeded, it means the remaining 90% stopped at pilots (small-scale trial applications) or failed to settle into operations. There are three major barriers. First, in a "no steady state" environment where data and processes keep changing, models quickly become obsolete. Second, the "Data Island" structure, in which each factory has different data schemas and shop-floor practices. Third, the problem of "hidden variables," where defect data and critical variables are not recorded. Additionally, an approach focused on "point AI technology verification (PoC)"—which validates technical possibilities of AI only in certain functions and situations separated from field operations—also lowers the probability of success. To raise the success rate of Physical AI, one must first prepare "operational automation capabilities" rather than focus on the technology itself. It is essential to standardize core in-factory processes (production, quality, equipment, logistics) and rebuild the data and workflow among sensors, equipment, and information technology (IT) systems into a single process through "business process management (BPM)." You also need a system that enables flexible experimentation with process changes and new product introductions through a Digital Twin. A "joint operations organization" in which field organizations share outcomes together is also important."
You are leading the research for a national Physical AI project worth around 1 trillion won.
The "Physical AI core technology demonstration (Jeonbuk specialized model development)" project is closer to a national experiment to turn a specific region into a Physical AI living lab than a simple research task. KAIST designs fundamental technologies and platforms under the theme of collaborative intelligence Physical AI, while Jeonbuk National University, local governments, and corporations verify the technologies in real factory, robot, and logistics environments. In that it attempts to create a regional-level Physical AI industrial ecosystem model beyond individual corporations' PoCs, it will be an important turning point for the development of Physical AI in Korea."
Your message to corporations and the government as the Physical AI era arrives.
We must break away from the defeatism that, because we fell behind in the LLM race, we should give up on AI. Korea was never a country that only does software. It is a country strong in "things that move," such as factories, equipment, automobiles, shipbuilding, and infrastructure. Physical AI is a game changer that can leverage this very strength. Because Physical AI is difficult to succeed through the efforts of individual corporations alone, a national-level platform strategy is needed in which the government and local governments, large, mid-sized, and small corporations, and universities and research institutes participate together. Now is not the time to say "let's study the technology more," but the time to "start experimenting in factories and cities, even on a small scale." Physical AI is not a theory on a desk, but a technology that matures through trial and error in the field. We must boldly increase test beds and demonstration projects on the assumption of failure, and build the accumulated experience there into a shared national asset."