KAIST develops AI that learns human judgment from a handful of videos

South Korean researchers have secured a core foundational technology that overcomes one of the key obstacles to the physical AI era: the limitations of data collection.

The breakthrough is expected to dramatically reduce the data-building costs and testing time companies face when developing new robots or autonomous driving systems.

KAIST announced Wednesday that a research team led by Professor Yoo Chang-dong of the Department of Electrical Engineering has developed a new technology called VOTP (Video-based Optimal TransPort Preference). The system enables AI to learn human intent and judgment criteria from just a few preference videos, rather than thousands or tens of thousands of human evaluation data points.

AI technology has rapidly evolved beyond generative AI that writes text and creates images, moving into the era of physical AI — systems that operate real machinery and act in the physical world. Representative examples include robots that perform dangerous tasks in factories, self-driving cars that assess road conditions on their own, and medical robots that carry out precision surgery.

Realizing the practical potential of physical AI, however, has required clearing a significant hurdle: teaching machines to evaluate whether their actions align with human intent and to judge which behaviors are more desirable — in other words, learning human-level evaluation standards.

The research team drew inspiration from the way people learn new tasks after watching only a few demonstrations. VOTP helps AI independently identify the behavior patterns humans prefer by using just a small number of videos showing good and bad examples. Rather than requiring people to manually evaluate vast amounts of data, the system allows AI to understand human judgment criteria and generalize that understanding across a wide range of situations.

The central idea behind the research is that intelligent machines such as robots and self-driving cars can quickly grasp human intent from only a small number of videos encoding human preferences.

"The technology can be applied broadly — from robotic arm control, humanoid robots, self-driving cars, smart factories and drones to surgical robots and even AI agents that directly operate computers," Professor Yoo said. "It can serve as a core foundational technology for all physical AI systems that need to learn human intent and satisfaction."

The findings were selected for presentation at ICML (International Conference on Machine Learning) 2026, the world's most prestigious AI research conference.