Korea is a global powerhouse in manufacturing robotics. According to the International Federation of Robotics (IFR), Korea's manufacturing robot density stands at 1,220 units per 10,000 workers — the highest in the world and roughly six times the global average of 177. Yet having the most robots does not guarantee top-tier competitiveness in physical AI. The country critically lacks the data needed to make robots perform useful real-world tasks.
Industry sources said Thursday that Korea's accumulation of action-based training data remains at an early stage compared with global leaders. "Action-based training data built in Korea currently stands at less than 10% of global levels," said Yoo Tae-jun, chairman of the Korea Physical AI Association. "We have not yet reached a stage where we can build industrial competitiveness." He added, "What matters most is not sheer volume but a data structure that links context, judgment and action — and that is our weakest point."
Action data cannot be scraped from the internet in bulk like text. It is accumulated over long periods as robots directly observe human actions through video or in-person demonstrations, imitate them and practice repeatedly. Without real-world data, securing physical AI competitiveness is difficult. In this regard, Korea faces an uphill battle.
For physical AI models to carry out actual tasks, they need scenario-type data in which "context-judgment-action" flows as a single sequence. A robot must recognize an object in a given process, record what judgment it made, what motion it executed and whether the outcome was a success or failure — all linked in one dataset. However, most field data in Korea is fragmented by individual equipment, with surrounding context severed, making it difficult to feed directly into AI training.
The biggest reason for this limitation is that manufacturing data is mostly locked inside companies and has not been standardized. "Each robot uses different cameras, pressure sensors and joint structures, and the data collection formats, intervals and resolutions all differ," said Kim Yu-cheol, head of strategy at LG AI Research. "Even when data is gathered, it cannot be used for training as-is." This means standardization work to consolidate data from various formats into a trainable form is essential. Kim Ki-hun, CEO of Movensys, also noted, "In high-speed processes measured in milliseconds, such as secondary battery manufacturing, hundreds of thousands of data points must be processed in real time — which is not easy with current infrastructure."
