Physical AI technology that accurately recognizes transparent objects such as glass or water and predicts future situations (AI-generated image). [Provided by KAIST]
Physical AI technology that accurately recognizes transparent objects such as glass or water and predicts future situations (AI-generated image). [Provided by KAIST]

South Korean researchers have secured key physical AI technologies that could accelerate the development of next-generation autonomous systems operating in real-world environments, including self-driving cars, humanoid robots, industrial robots and digital twins.

KAIST announced Sunday that a research team led by Yoon Sung-eui, a professor in the School of Computing, has developed physical AI technologies capable of accurately recognizing transparent objects such as glass or water and predicting future situations.

The research is significant for linking visual recognition, physical understanding, future prediction and action planning into a single technological pipeline. The work lays the groundwork for AI to carry out the full cycle of perceiving, understanding, predicting and acting — a foundation expected to broaden the performance and application range of various autonomous systems.

The team developed GLINT (transparent-environment visual recognition technology), enabling AI to accurately identify transparent objects such as glass.

By separately analyzing reflections on glass surfaces and objects visible behind them, the researchers enabled AI to correctly interpret scenes even in transparent environments.

The team also developed RadioGS (scene reconstruction technology for understanding light and material properties), which enables AI to understand how light strikes an object, reflects off it and scatters.

The same object can look different under sunlight versus indoor lighting, and conventional AI has been susceptible to such variations. The team trained AI to learn the interaction between light and objects, allowing it to more accurately understand material properties and surroundings regardless of lighting conditions.

The KAIST research team behind the study. From left: Andrew Jeong, Na Yeong-ju, Lee Ju-min, Lee Se-bin (doctoral students), professor Yoon Sung-eui and Han Gyu-beom (doctoral student). [Provided by KAIST]
The KAIST research team behind the study. From left: Andrew Jeong, Na Yeong-ju, Lee Ju-min, Lee Se-bin (doctoral students), professor Yoon Sung-eui and Han Gyu-beom (doctoral student). [Provided by KAIST]

The team developed Visual-RRT (image-based robot path-planning technology), which bridges visual information and real-world action.

In robot experiments, the system successfully guided a robot to its destination using only a single photograph, demonstrating its potential across fields including service robots and autonomous driving robots.

The team also developed CLaD (future-prediction-based action planning technology), which enables AI to anticipate future situations and plan the most appropriate course of action before acting.

The technologies can be applied across a wide range of fields requiring real-world understanding and decision-making, including self-driving cars, service robots, industrial robots, digital twins, augmented reality, mixed reality, smart manufacturing and logistics automation.

"Through this research, AI will be able to move beyond simple recognition systems toward understanding the real world, predicting the future and taking action," Yoon said. "We hope this contributes to the advancement of various physical AI technologies operating in real-world environments, including self-driving cars and humanoid robots."

The findings were presented at ICLR 2026 and CVPR 2026, two of the world's most prestigious AI conferences.


nbgkoo@heraldcorp.com