ETRI receives global recognition for its AI technology for recognizing unstructured outdoor environments

Korean researchers won second place at the Semantic Segmentation International Challenge at ICRA 2025, the world’s most prestigious robotics conference. This raises expectations for the researchers’ performance in the field of future robot eyes that see through images.

Electronics and Telecommunications Research Institute (ETRI) has announced that it participated in the Goose 2D Semantic Segmentation Challenge at the Field Robotics Workshop of ICRA 2025 and achieved excellent results. This achievement once again proves the international competitiveness of Korea’s AI-based image recognition technology.

This challenge tests the ability of artificial intelligence technology to accurately classify objects such as bushes, stones, trees, and the ground, pixel by pixel, based on 2D image data captured in unstructured terrain such as fields, forests, and construction sites that field robots actually encounter.

Participants develop pre-trained models based on limited training data and then compete on the accuracy of their semantic segmentation against an undisclosed test dataset. The higher the mIoU (mean Intersection over Union), the better the AI model's ability to recognize objects in unstructured terrain, such as forests and construction sites.

A total of eight teams from South Korea, Germany, and Taiwan participated in the challenge. ETRI participated as a team called the ‘Mobility AI Group’ formed by Principal Researcher Ahn Soo Yong and Postdoctoral Researcher Kim Won Jun from the Mobility AI Convergence Research Section (Director Lee Lae Kyoung) at Daegu-Gyeongbuk Research Division. The researchers implemented an artificial intelligence model that performs well in complex, unstructured outdoor environments, resulting in a second-place finish in the competition.

ETRI’s skill has been recognized internationally for implementing artificial intelligence models that can segment objects with high accuracy in unstructured outdoor environments. This also means that robots have acquired the ability to distinguish and understand objects more clearly from a human perspective. The top place in the challenge went to a team from Taiwan’s National Cheng Kung University.

The dataset provided in this challenge is fundamentally different from traditional city-based semantic segmentation datasets. Unlike typical datasets that consist of structured roads and clear objects, the dataset for this challenge consists of footage captured in unstructured outdoor environments such as forests, fields, and undeveloped areas.

This includes realistic environmental elements such as light variations, irregular structures, and visual obstructions, and the visual similarity between objects is also an important factor in testing the model’s ability to generalize. These characteristics serve as very high standards for evaluating the technological maturity of practical field robots.

Byun Woo Jin, vice president of the Daegu-Gyeongbuk Research Division, said, “This award is of great significance as it is internationally recognized by the world’s most prestigious robotics society for our technology and research achievements. In particular, image-based semantic segmentation technology can be used for a wide range of mobility applications, including autonomous driving, logistics, and industrial robotics.”

ETRI’s Daegu-Gyeongbuk Research Division focuses on the development of practical technologies that can be applied in the field, and has expressed its ambition to become a research institute that contributes substantially to the domestic and global industrial ecosystem.

Building on this achievement, the researchers are now focusing on advancing autonomous driving technology in unstructured terrain. They plan to continue developing robust semantic recognition-based object segmentation technology to enable stable operation in forests, farmland, construction sites, disaster response areas, and other areas where it is difficult to secure recognition accuracy and driving stability with existing autonomous driving technologies. Through this, they are actively promoting the expansion of the technology into various applications such as industrial robots, agricultural robots, and disaster response robots.

The award is seen as more than just a competition achievement, but as a real milestone that provides international validation of the technology’s applicability in a real-world field environment. ETRI will continue to lead innovation in various industries centered on AI-based image cognitive technology.

###

This achievement was carried out as part of the ‘ICT Convergence Technology Advanced Support Project (Mobility) based on the Daegu-Gyeongbuk Regional Industry’ of the Ministry of Science and ICT.