"In Korea, there's a culture of pouring soju into someone's glass when it's empty. What would it take for a robot to do this?"
Daniel Lee, a professor at Cornell University, posed this question on the 28th as he addressed the challenges of commercializing physical AI during the session "Redesigning Industry: A Physical Revolution Led by Robotics" at "Seoul Forum 2026," held under the theme "Beyond Intelligence, Toward a New Engine of Industry" at the Shilla Hotel in Jung-gu, Seoul.
His point was that actions trivial for humans—such as picking up a soju bottle and pouring a drink into a glass—are formidable hurdles for robots. He raised the so-called "Moravec's Paradox" as the core challenge—the ironic phenomenon in which tasks easy for robots are difficult for humans, and tasks easy for humans are difficult for robots.
"The reality is that even a simple task like a robot lifting a soju bottle and pouring a drink into a glass is tricky," Professor Lee explained. "It's an obvious intuition for humans, but for a robot to accomplish it, it must understand physical realities such as the bottle's coefficient of friction and how to keep the drink from spilling onto the other person if it slips."
Other experts also raised the issue of data learning. Song Ki-young, CEO of Holiday Robotics, a domestic startup developing humanoid robots, said, "For humanoid robots to be deployed on manufacturing sites, they must achieve task accuracy of around 99.9%, but AI technology capable of realizing this has not yet emerged." He added, "In particular, because the cost of collecting data is high, we are focusing on reinforcement learning that uses simulations incorporating virtual data rather than actual data."
Strategies for surviving the global AI hegemony competition were also presented. Director Jang said, "The government and industry are currently joining hands and have been pursuing the development of a 'World Foundation Model' since last year, and we are also carefully preparing in the physical AI field." He added, "If industry responds nimbly on an R&D basis in line with the government's development roadmap, we can lead the market with our own independent roadmap that doesn't fall behind the U.S. and China."
There was also a suggestion that technology allowing robots to intuitively understand reality—like humans, without prior learning—is needed. Professor Lee, who defined the current level of AI technology as a "black box," presented "white box" theory, which clarifies AI's complex computational processes, as a solution. Unlike a black box, which infers causes from results, a white box is a concept that explains how inputs are processed to reach conclusions.
"We need to open up AI's black box and examine how AI converts data structures into specific tasks," Professor Lee emphasized. "Observing the geometric features of an AI system's computational operations can contribute to advancing AI models." He added, "Understanding and interpreting what happens inside AI will guarantee AI's safety and performance in actual physical environments."
There was also a forecast that humanoid robot technology will follow the path of the autonomous driving industry. Hong Sung-soo, a professor at Seoul National University, predicted, "In the autonomous driving field, end-to-end technology that handles everything from a vehicle's surroundings sensing to control emerged in 2023, and within two years it became widespread, making it possible to identify the specific causes of car accidents." He added, "The two years of progress in autonomous driving technology is comparable to the past 60 years, and humanoid robots will follow a similar path."
