Korea Tops World in Robot Density but Trails in Physical AI Data

The world's leading physical artificial intelligence (AI) companies have already entered a fierce race to secure large-scale real-world behavioral data. U.S. startup Physical Intelligence said it used 10,000 hours of robot demonstration data to train its "π0 (pi-zero)" model, unveiled last year. Another Silicon Valley startup, GeneralistAI, announced that its physical AI model "GEN-0" was trained on 270,000 hours of real-world operational data. Industry observers project that the cumulative training data held by major algorithm companies will surpass one million hours by next year. Behavioral data cannot be scraped from the internet in bulk like text. It accumulates over long periods as robots directly observe human actions through video or in-person demonstrations, imitate them, and repeatedly practice — with each attempt recorded. Without real-world data, building competitiveness in physical AI is extremely difficult. In this regard, Korea's position is far from comfortable. According to industry sources on Monday, Korea's accumulation of behavior-based training data remains at an early stage compared to global leaders. "Behavior-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 data structures that connect context, judgment, and action — and that is our weakest link." Korea is a global manufacturing robotics powerhouse. 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, more than six times the global average of 177. In terms of factory automation centered on semiconductors, automobiles, and electronics, Korea already ranks among the top tier. However, these deployed robots do not automatically translate into humanoids or general-purpose intelligent robots for the physical AI era. For a physical AI model to perform actual tasks, it needs scenario-type data in which "context-judgment-action" are linked as one sequence. The data must capture what object a robot recognized during a given process, what judgment it made, what motion it executed, and whether the outcome was a success or failure — all connected in a single dataset. Yet most on-site data in Korea is fragmented by individual equipment, with contextual continuity severed, making it difficult to feed directly into AI training. The biggest cause of this limitation is that manufacturing data is mostly accumulated in closed silos within companies, with no standardization. "Each robot uses different cameras, pressure sensors, and joint structures, and the data collection formats, intervals, and resolutions all vary," 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 — consolidating data from diverse formats into a trainable form — is essential. Kim Ki-hun, CEO of Mobensys, 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." Converting tacit knowledge into data is another critical challenge. This involves transforming the hand skills, intuition, experience, and split-second judgment of skilled workers — know-how that is difficult to explain in words or manuals — into forms that can be observed, measured, and recorded. However, companies are reluctant to release such data externally for fear of leaking process know-how. Many skilled workers also resist cooperation, anxious that digitizing their techniques and judgment could ultimately be used to replace them. The problem is that the aging population is accelerating the departure of skilled workers from factory floors. Processes that depend on the senses and experience of veterans — such as precision assembly, welding, inspection, and sorting — are highly likely to be lost forever if not captured as data now. "Converting tacit knowledge into data should be approached not as a mere technical issue but as a matter of preserving industrial knowledge," an industry official stressed. "This may be our last chance." Still, simply piling up data is not enough. Voices within the industry increasingly emphasize the need to go beyond data collection and focus on processing and verifying it into high-quality datasets. Experts say Korea must standardize data from different robots and sensors into trainable formats and concentrate on accumulating high-quality data in areas where the country has unique strengths. "Korea has advantages in manufacturing, quality control, precision processes, and industrial diversity," Yoo said. "If we fail to set a direction within two to three years, the gap is likely to become permanent. We must compete on data quality, not quantity."

Korea Tops World in Robot Density but Trails in Physical AI Data

Related Video

An Islamic NATO

Token Fines and Designed Traps: Holding Platforms Accountable for Youth Addiction

Editorial: Non-Business Property Crackdown Must Not Chill Corporate Investment

President Lee Says Fixed-Term Employment Law Has Become 'Hiring Ban'

Editorial: Performance-Driven Microfinance Must Not Ignore Growing Defaults

Listed Firms Hoard Record 107 Trillion Won in Non-Business Real Estate

Listed REITs Market Cap Tops 10 Trillion Won, Milestone After 25 Years

SKT Partners with Arm and Rebellions to Develop GPU-Alternative AI Server Solution

Anthropic Considers In-House Chip Development as AI Big Tech Accelerates Exit from Nvidia

53% of Gen Z Say AI Has Made Cover Letters Obsolete in Hiring