Korea-US research team wins ICRA robot vision challenge

A joint Korean-American research team has taken first place at the world's most prestigious robotics conference, defeating 56 competing teams from around the globe.

The Daegu Gyeongbuk Institute of Science and Technology (DGIST) announced Tuesday that a team led by Professor Yoon Sung-hoon of its Department of Electrical Engineering and Computer Science, together with postdoctoral researcher Im Hyung-tae of the Massachusetts Institute of Technology (MIT), won the GOOSE 2D Semantic Segmentation Challenge at the Field Robotics Workshop of the 2026 International Conference on Robotics and Automation (ICRA).

Co-organized by Germany's Fraunhofer IOSB, the Bundeswehr University Munich and the University of Koblenz, the challenge evaluates how precisely field robots can interpret complex scenes in unstructured real-world environments. Unlike conventional autonomous driving datasets collected mainly on well-organized urban roads, the GOOSE dataset used in the competition draws on field robot data gathered in unpredictable, unstructured outdoor settings.

That dataset covers unstructured outdoor data from multiple platforms, including excavators and quadruped robots, making it considerably more demanding than standard urban road environments. This year's competition raised the bar further by expanding the evaluation criteria to 64 detailed classes, requiring teams to accurately identify even "rare objects" that appear with extremely low frequency in the field.

The joint research team developed an original framework that combines Meta's latest self-supervised foundation model, DINOv3, with the image segmentation model Mask2Former. The system delivered stable visual recognition across a range of challenging conditions — varying light levels, irregular terrain and complex backgrounds — in both urban environments and rugged, unstructured settings such as forests.

The team also dramatically reduced critical recognition failures that could lead to accidents by maximizing detection of rare objects that AI systems tend to miss due to limited training data, improving overall safety. The technology is expected to find broad application beyond autonomous vehicles, extending to field robotics in disaster response, smart agriculture, construction sites and other industries.

"This achievement demonstrates the excellence of AI-based image recognition technology on the world stage and proves the feasibility of physical AI," Professor Yoon said.