"There is demand to introduce and utilize physical artificial intelligence (AI) in elderly care and caregiving settings. To discover diverse demand and accumulate data, the fragmented ecosystem must interact with one another."
Choi Hong-seop, CEO of Maum AI, described the reality of Korea's behavioral data ecosystem at a Physical AI Demonstration Lab open briefing held at KAIST in Daejeon on May 29. He explained that a strategy of first building training data in fields with relatively clear demand — such as caregiving, logistics, and manufacturing — should take priority over spaces like private homes, where resistance to data collection is strong.
In Korea, efforts by private companies to accumulate physical AI behavioral data have recently taken concrete shape. Companies are directly seeking out sites with clear demand in caregiving, logistics, manufacturing, and services to secure data. As one example, Config Intelligence Korea has built a human data factory in Hanoi, Vietnam, producing behavioral data tailored to the work environments its clients require. For household tasks, workers repeatedly perform actions such as pouring water or folding towels; for industrial settings, they carry and manipulate objects to build training datasets. The company is reported to have collected more than 100,000 hours of data since beginning full-scale operations in the second half of last year. Realworld is conducting demonstrations in manufacturing, logistics, and services in partnership with CJ, Lotte Hotel, SK, and Japan's KDDI. Its distinguishing feature is the use of multiple cameras and wearable sensors to directly capture worker behavioral data generated at actual sites, rather than relying on simple simulations.
However, policy support is essential for these private-sector efforts to produce results. The government has also recognized the next three years as the golden window that will determine dominance in physical AI and has unveiled its "Strategy for Securing Core Competitiveness in Physical AI."
The government plans to develop a general-purpose robot foundation model capable of autonomously planning and performing precision tasks over extended periods, much like a human. To ensure the model operates in the real world without errors or delays, the government will build a "world model" that generates large-scale training data and supports virtual experiments, and will also develop a computing platform based on high-performance, low-power, low-latency AI semiconductors.
Furthermore, the government plans to identify fields with high demand for physical AI applications so that technologies can be quickly demonstrated and scaled to meet private-sector needs. The plan envisions immediately introducing and demonstrating physical AI technologies in areas closely tied to daily life — including logistics, agriculture, disaster response and safety, and caregiving and household services — to deliver tangible results within one to two years. Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon said, "The next three years are a golden window during which Korea must mobilize all national capabilities to leap forward as a physical AI powerhouse," adding, "We must rally the strengths of industry, academia, and research institutes to operate a full-cycle support system spanning from independent physical AI technology development to on-site deployment."
Industry voices also call for connectors that can bring the ecosystem together — linking demand sources with supply companies. The point is that rather than collecting data in one place for the physical AI data race, the nation needs a systematic framework for gathering and sharing data at the national level. "The physical AI industry is a large ecosystem, so interaction among companies is necessary," CEO Choi said. "I hope a structure emerges in which larger alliances are formed and companies compete within them."