DGIST-MIT Research Team Wins First Place at Global Robot Vision Challenge, Demonstrates Unstructured Environment Recognition Technology

Surpassing 56 Teams in the ICRA 2026 GOOSE Challenge
Demonstrating Core Visual Perception Technology for Autonomous Vehicles and Disaster Response Robots

A joint research team from Daegu Gyeongbuk Institute of Science and Technology (DGIST) and the Massachusetts Institute of Technology (MIT) has secured first place in a robot vision challenge at the world’s most prestigious robotics conference. By demonstrating artificial intelligence (AI) technology capable of accurately recognizing rare objects even in unstructured outdoor environments, the team has proven its competitiveness in the next-generation autonomous driving and field robotics sectors.

On June 9, DGIST announced that the joint research team led by Professor Seong-Hun Yoon from the Department of Electrical Engineering and Computer Science at DGIST and Postdoctoral Researcher Hyungtae Lim from MIT achieved first place among 56 global teams in the 'GOOSE 2D Semantic Segmentation Challenge' at the Field Robotics Workshop during the 2026 International Conference on Robotics and Automation (ICRA 2026).

Professor Sung-Hoon Yoon's team at DGIST won first place in the 'GOOSE 2D Semantic Segmentation Challenge' category at '2026 ICRA'. Provided by DGIST

This competition was co-hosted by the Fraunhofer IOSB Institute in Germany, Bundeswehr University Munich, and University of Koblenz. It is an international contest that evaluates how accurately field robots comprehend complex scenes encountered in real outdoor environments.

The 'GOOSE dataset' used in the competition is based on unstructured outdoor data collected from various platforms, including excavators and quadruped robots. Unlike typical autonomous driving datasets that focus on urban roads, this dataset reflects real-world environments with irregular terrain and diverse obstacles, thus presenting a higher level of difficulty.

Notably, this year’s challenge expanded the evaluation to 64 detailed classes, requiring precise identification of 'long-tailed classes'—rare objects with extremely low occurrence rates. Failure to identify these rare objects can lead to safety accidents during actual autonomous driving or field robot operations, making this a key assessment factor.

The research team developed a proprietary framework that combines 'DINOv3,' a self-supervised foundation model developed by the American AI company Meta, with the image segmentation model 'Mask2Former.' This technology demonstrated stable visual recognition performance under various environmental changes, including fluctuations in lighting, complex backgrounds, and unstructured terrain.

Research team led by Professor Seunghun Yoon of the Department of Electrical Engineering and Computer Science at DGIST. From left: Professor Yoon, Sangjin Lee, Hyobin Choi, Jaeil Park. Provided by DGIST

In particular, the team's framework significantly improved the detection performance of rare objects, which AI models often overlook due to data scarcity, thus successfully reducing critical recognition errors. As a result, this technology is expected to be applicable not only to autonomous vehicles but also to a wide range of industries, including disaster response robots, smart agriculture, and construction site robots.

Professor Seong-Hun Yoon of the Department of Electrical Engineering and Computer Science at DGIST stated, "Technology that can accurately interpret scenes in unpredictable, unstructured outdoor environments is fundamental to ensuring the autonomy and safety of field robots. Building on this achievement, we will continue to advance research on robust visual perception technology that can be immediately applied to real-world industrial settings."

DGIST emphasized that this achievement was made possible through joint research with international institutions such as MIT, and expressed expectations that it will contribute to expanding global cooperation in robotics and AI research, as well as to the commercialization of these technologies.

This content was produced with the assistance of AI translation services.