It Would Take 5 Years to Collect 10,000 Hours of Robot Data
The biggest obstacle facing physical AI is a shortage of data. AI robots that operate in the real world—the so-called physical AI industry—are being held back by the very resource that underpins AI model training. Without data to learn from, robots cannot develop the ability to make their own judgments and act on them, and the industry is searching for new ways to collect it.
The way physical AI acquires data is fundamentally different from large language models (LLMs). While LLMs like ChatGPT can train on text and video publicly available on the internet, the physical data robots need cannot be obtained online and must be gathered by hand. Cameras are strapped to workers' chests and foreheads, and sensors attached to their hands, recording their movements and work processes on a computer. According to the Ministry of Science and ICT, physical AI requires 10 to 100 times more data and GPU computing resources than LLMs. Song Chang-jong, head of the ministry's Device AX Innovation Team, explained, "Collecting 10,000 hours' worth of data would take five full years at eight hours a day." The shortfall can be made up with synthetic data. Real and synthetic data are currently used in equal measure, and the government expects synthetic data's share to grow to as much as 90%.
An even bigger problem is the lack of sites where data can be collected and robots trained. AI companies sign nondisclosure agreements and collect data or run demonstrations at individual workplaces, but the space available for such use is limited. For workers, wearing filming equipment on the head and chest is a hassle, and the prospect of robots being trained on data from their own work can feel burdensome. Government agencies are in no better position. The science ministry sounded out several private logistics companies about running demonstration projects, but the firms were reluctant to open their facilities, so the ministry decided to proceed with the projects through the Korea Post headquarters instead.
China moved early to accumulate data. According to materials released by the office of Democratic Party lawmaker Park Hee-seung, the number of physical AI training grounds in China expanded from 24 at the end of 2025 to more than 70 by late June this year. Roughly 140 billion won was invested in Beijing's Shijingshan District data training center alone. Korea has no dedicated training infrastructure on that scale.
The science ministry will launch a medium- to long-term physical AI demonstration project next year, to run for two years. Korea Post is also conducting physical AI demonstrations with the Ministry of Land, Infrastructure and Transport, and is considering installing loading-and-unloading robot arms at some mail processing centers around the end of this year to deploy them during peak volume periods.
