KAIST researchers develop method for AI to learn human preferences from videos

Researchers at KAIST in South Korea have developed a technique that allows machines to learn human preferences and evaluation criteria from a limited number of videos. The approach, named Video-based Optimal TransPort Preference (VOTP), was prepared under the leadership of Prof. Chang Deok-yu of the Department of Electrical Engineering. The method is designed to replace systems that previously required tens of thousands of manually labeled human evaluations. Instead, it relies on analyzing a small number of video examples containing both successful and unsuccessful outcomes, allowing systems such as surgical suturing robots or autonomous vehicles navigating busy intersections to align their behavior with human expectations without relying on rigid rules. In experiments, the technology achieved successful results across different tasks and conditions and was able to transfer what it had learned to situations it had not previously encountered. Researchers said VOTP can lower data collection costs while supporting a broad range of applications, including factory automation, humanoid robots, unmanned aerial vehicles and advanced surgical systems.