New VOTP technology lets AI learn human judgment from just a handful of preference videos, dramatically cutting development costs for robots and self-driving cars

Professor Yoo Chang-dong (left) and his research team at KAIST, who conducted the study. [KAIST]
Professor Yoo Chang-dong (left) and his research team at KAIST, who conducted the study. [KAIST]

South Korean researchers have secured a foundational technology that overcomes one of the biggest obstacles standing between today's AI and the physical AI era: the limits of data collection.

The breakthrough is expected to dramatically cut the enormous 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 technology enables AI to learn human intent and judgment criteria from just a handful of preference videos, rather than thousands or tens of thousands of human evaluation data points.

AI technology has been rapidly evolving beyond generative AI that writes text and creates images, moving toward "physical AI" that operates real machinery and acts in the physical world. Prominent examples include robots that perform dangerous tasks in factories, self-driving cars that assess road conditions on their own, and surgical robots that carry out precise medical procedures.

Yet realizing physical AI in practice has required clearing a critical 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 assessment standards.

The research team drew inspiration from the way people can pick up new tasks after watching just a few demonstrations. VOTP uses a small number of videos showing good and bad examples to help AI independently identify the behavioral patterns humans prefer. Rather than requiring people to manually evaluate vast quantities of data, the system allows AI to grasp human judgment criteria and generalize that understanding across a wide range of situations.

A research image illustrating the VOTP technology (AI-generated). [KAIST]
A research image illustrating the VOTP technology (AI-generated). [KAIST]

The core insight 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 algorithm developed to achieve this has demonstrated its effectiveness and generalization across diverse environments and tasks through extensive experiments.

This approach can significantly reduce the human feedback and data-building costs required for physical AI development. Because robots, self-driving cars, and industrial machines can learn behavior that meets human expectations from just a few examples, the technology is expected to dramatically shorten development timelines and lower costs.

"The technology can be applied broadly — from robotic arm control, humanoid robots, self-driving cars, smart factories, drones, and surgical robots to AI agents that directly operate computers," Professor Yoo said. "In particular, it can serve as a core foundational technology for any physical AI system that needs 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 academic conference.


nbgkoo@heraldcorp.com