Dennis Hong: Success of Physical AI Depends on Real-World Data

"While generative AI data is available on the internet, crucial data for physical AI is found in the field. By deploying robots in actual industrial settings to generate high-quality physical data, we can create a 'manufacturing-centric virtuous cycle ecosystem' that uniquely positions Korea in the physical AI landscape."

Dennis Hong, a professor of mechanical and aerospace engineering at the University of California, Los Angeles (UCLA) and director of the Robotics and Mechanisms Laboratory (RoMeLa), presented this perspective during his keynote speech at the 18th Good Growth, Good Jobs Global Forum (2026 GGGF) held at The Plaza Hotel in Jung-gu, Seoul, on September 2.

Hong emphasized that Korea does not need to simply replicate the U.S. model of AI development centered around big tech or the Chinese approach that leverages a powerful hardware supply chain. He noted that Korea possesses world-class manufacturing sectors in automotive, semiconductors, and shipbuilding, and should leverage its strengths in field data to enhance industrial competitiveness.

He identified a significant shift in robotics over the past couple of years, moving from traditional control mechanisms that required manual coding of algorithms based on sensor inputs to a more integrated approach where a single AI model handles everything from sensor input to final motion control.

However, he pointed out that implementing this AI-based control method requires substantial data. Unlike generative AI or large language models, the physical AI sector faces a severe 'data drought.' While text, images, and videos can be amassed from the internet for training, precise physical data such as joint positions, speeds, accelerations, and friction coefficients necessary for training robots are not readily available online.

Current data collection methods employed by global tech companies also show limitations. For instance, virtual world simulation training often leads to errors when applied to real robots due to discrepancies with reality. Collecting data through virtual reality (VR) equipment is time-consuming and costly, and video-based learning from platforms like YouTube fails to provide tactile data such as force magnitude or contact sensations.

Hong stressed that Korea's unique manufacturing environment could serve as a crucial asset in the era of physical AI, as its top-tier K-manufacturing factories can extract the most sophisticated and extensive real-time physical data. He urged caution against the excessive hype surrounding the robotics industry, noting that while AI-based control methods offer remarkable flexibility, they also present a 'black box' problem that complicates fault diagnosis when errors occur. He advocated for advanced research that combines traditional model-based control methods, which have evolved over the past century, with AI approaches.

"The success of humanoid and physical AI ultimately hinges on who can most efficiently collect and apply precise real-world data," Hong stated. He called for the Korean industrial sector to maximize its substantial strengths in manufacturing to establish a practical physical AI ecosystem.