GIST Develops 'Robot Vision AI' for Untrained Object Recognition

Research team led by Professor Lee Gyubin, School of Convergence Technology, GIST. Photo by GIST

[Asia Economy Honam Reporting Headquarters Reporter Cho Hyung-joo] The research team led by Professor Lee Gyu-bin of the Convergence Technology Interdisciplinary Department at GIST (Gwangju Institute of Science and Technology, President Kim Ki-seon) announced on the 16th that they have developed a deep learning technology that simultaneously detects the visible area, occluded area, and occlusion status of unseen objects even in complex robotic environments through hierarchical occlusion modeling.

For a robot to manipulate objects in a new environment, it is necessary to accurately detect new objects that have not been previously learned.

Instance segmentation, which detects object-specific regions from images, is a core research area in deep learning and robot vision, and various studies have been proposed. However, there is a limitation in that only objects of pre-learned categories can be recognized or only the visible areas of unseen objects can be detected.

To address this issue, the research team proposed a new task called amodal instance segmentation of unseen objects, which aims to simultaneously detect not only the visible areas of unseen objects but also the occluded areas and occlusion status.

Additionally, the team proposed hierarchical occlusion modeling to effectively consider occlusion relationships between objects and released new virtual and real environment datasets for training and evaluation.

The algorithm proposed by the research team achieved world-class performance on three datasets, confirming that it can significantly improve robotic recognition performance in complex environments.

They also confirmed that it can be utilized in various real robot tasks by enabling robots to grasp occluded target objects.

Professor Lee Gyu-bin said, "Through this research, we confirmed that even when new objects are given in complex unstructured environments, it is possible to recognize not only the visible areas of objects but also the occluded areas," adding, "Unseen object recognition is expected to be a core technology in the robotics field for applying robots in various environments such as factories and homes."

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