"Artificial intelligence (AI) may be far superior to humans in highly complex domains, but it can appear absurdly foolish in everyday tasks that humans perform with ease."
Daniel Lee, a globally recognized authority on AI robotics and professor at Cornell University's College of Engineering, addressed the urgent challenges facing physical AI during the "Industrial Redesign: The Physical Revolution Led by Robotics" session at Seoul Forum 2026, held Tuesday at the Shilla Hotel in Jung-gu, Seoul, under the theme "Beyond Intelligence, Toward a New Engine of Industry." Lee said humanoid robots embodying physical AI are running into what is known as "Moravec's paradox."
Lee, who has researched AI fields including robotics, machine learning and neuroscience, acknowledged the limits of current robotics technology by describing concrete situations. "Some say AI technology is approaching superintelligence, but in reality, even the task of a robot picking up a soju bottle, pouring it into a glass and handing it to a person is challenging," he said. "It is easy for a human, but for a robot to actually do this, it must understand the friction coefficient of the bottle, and what to do if the bottle slips when grasped — how to ensure the liquor spills onto the floor rather than onto the other person."
For this reason, the "world model" — a neural network that understands the dynamics of the physical world — is rapidly emerging as an alternative for commercializing physical AI. "Large language models (LLMs) are based on text, so they have limits in understanding physical reality," Lee said. "Beyond improving hardware performance such as precision, it is becoming important for robots to learn to understand the external world through world models."
His point was that robots need a technology that allows them to intuitively understand reality without being taught in advance, much like humans do. Defining the current level of AI technology as a "black box," Lee proposed the "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 explains how inputs are processed to reach conclusions.
"We must open up AI's black box and examine how AI converts data structures into specific tasks," Lee said. "Observing the geometric characteristics of AI systems' computational operations can contribute to advancing AI models." He added, "Understanding and interpreting what happens inside AI will guarantee the safety and performance of AI in actual physical environments."
In particular, he advised that Korea should focus on efficiency rather than quantitative AI development. "Korea's population is no longer growing, so it is reaching a limit in data volume," Lee stressed. "Rather than simply focusing on scaling up, intelligent construction of AI is needed." He added, "Korea lags behind the United States and China in terms of economies of scale, but its high level of technical talent is its competitive edge." In the global AI model competition, he said, Korea must leverage its talent as a strength to accelerate the completion of efficient AI.
Other panelists also offered alternatives during the session's discussion to supplement the real-world data-driven training approach. Song Ki-young, CEO of Holiday Robotics, a domestic humanoid robot development startup, said, "For humanoid robots to be deployed at manufacturing sites, they must achieve 99.9% task accuracy, but AI technology that can implement this has yet to emerge." He added, "Especially since data collection costs are high, we are focusing on reinforcement learning that simulates by incorporating virtual data rather than real data."
A diagnosis was also offered that vision language model (VLM) technology, which processes visual data such as images and videos, could contribute to the development of the humanoid industry. Hong Sung-soo, a professor at Seoul National University, said, "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 just two years it became widespread, allowing the causes of vehicle accidents to be identified in detail." He predicted, "The two years of progress in autonomous driving technology is comparable to the past 60 years, and humanoid robots will likely follow a similar path."
