The age of physical artisan AI
Kim Dae-shik The authos is a professor at KAIST. Geoffrey Hinton's 2012 research made “perceptual AI” possible. AI began to recognize the world. With the arrival of ChatGPT at the end of 2022, society has entered the era of “generative AI,” capable of understanding human language and producing information on demand. Yet perceptual and generative AI are only the beginning. Moving beyond simply generating what users want to see, so-called “agentic AI” can semi-autonomously carry out the actions and tasks users wish to perform. A screen reads 'AI in the physical world' as attendees gather during Rivian's first Autonomy and AI (artificial intelligence) Day, showcasing developments in self-driving technology, in Palo Alto, California, on Dec. 11, 2025. [REUTERS/YONHAP] Humans, however, live in an analog reality. Beyond actions in the digital realm, AI must also operate in the physical world, one with bodies, objects and tangible constraints. This is why technologies such as autonomous vehicles and humanoid robots, often grouped under “physical AI,” are seen as major blue oceans for future industries. Humanoid robots in particular are technologies Korea cannot afford to miss, given the continued importance of manufacturing. The ATLAS prototype robot by Boston Dynamics makes an appearance on stage during a press conference at the 2026 International CES, at the Mandalay Bay Convention Center in Las Vegas, Nevada, on Jan. 5, 2026. [UPI/YONHAP] Since humans settled into agriculture in the Middle East some 10,000 years ago, they have built houses, cities, factories and machines. All share one trait. They were designed for the human body. Homo sapiens grasp objects with two hands and move on two legs. As a result, most environments are built for the human form or require a humanlike body to operate and control them. In a world already designed around humans, a humanoid form is the most optimized way for machines to act on humanity’s behalf. The human body is precise and complex. A single hand is said to have 27 degrees of freedom. For decades, engineers attempted to manage such intricate motion through inverse kinematics. Yet the equations required to compute joint values for humanoid movement are extraordinarily complex, and perfect solutions are often impossible. This is why, until only a few years ago, humanoid robots struggled even to walk or perform simple actions. Related ArticleBoston Dynamics' market value likely exceeds $20 billion due to Atlas robotLG demonstrates CLOiD robot, aims for 'Zero Labor Home'At CES 2026, a humanoid robot showdown between Korea and ChinaHyundai's humanoid robot Atlas to debut at CES 2026 The challenges facing physical AI are similar to those encountered in other AI fields. For decades, scientists tried to solve object recognition with formulas. No matter how carefully they modeled and quantified the world, machines failed to truly recognize it. Only after the AI paradigm shifted from explanation to learning in 2012 did object recognition advance rapidly. Natural language processing followed a similar path. Even after being taught grammar and linguistic rules, machines could not understand human language. Once they began learning from hundreds of billions of sentences posted online, progress accelerated, culminating around 2022. ChatGPT can generate text, images and videos because it has learned from content uploaded to the internet over the past 30 years. For humanoid learning, however, data on human actions and joint movements barely exist. Developing physical AI, therefore, faces a high barrier to entry, requiring the creation of learning data itself. This picture, taken on November 5, 2025, shows Hiro Yamamoto, CEO of the company Enactic, teleoperating an OpenArm humanoid robotic arm at his office during an interview with AFP in Tokyo. A pair of swivelling, human-like robotic arms, built for physical artificial intelligence research, mirror the motions of an operator in a VR headset, twirling his hands like a magician. [AFP/YONHAP] There are two main ways to obtain behavioral data for humanoids. One is to generate it through computer simulation. Yet the analog world is astonishingly complex and unpredictable. Perfect motion in the physical world is unlikely to emerge solely from digital simulation. The alternative is to have humans demonstrate and teach actions directly. While this can precisely convey desired behaviors, it demands enormous time and effort. It also carries a critical limitation. A humanoid’s abilities will reflect the range and skill of the human teacher. Just as ChatGPT, trained only on English, would not understand Korean, a humanoid trained on the actions of a part-time worker or student is unlikely to assemble ships at a shipyard. Korea is unusual in that manufacturing still accounts for a large share of its economy. More than that, the spectrum is broad. During the pandemic, Europe was shocked to discover it had no mask factories, while Korea continued to produce everything from semiconductors and aircraft to paper straws and masks. This breadth may become a decisive strength in the era of physical AI. While ChatGPT is trained on publicly available internet data, physical AI trained on the movements of skilled workers in Korean factories could, in theory, be monopolized. As AI evolves rapidly from generative to agentic and physical forms, “physical artisan AI,” trained on the movements of masters across Korea, may one day shape the nation’s future. This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.
