[Abstract]

Reliable perception is a critical foundation for general-purpose robotic systems. In this talk, I will first present our recent research on advancing 3D LiDAR-based segmentation for robot navigation. Our effort focuses on maintaining high segmentation performance under adverse weather conditions, as well as mitigating the degradation caused by dynamic objects in complex environments. These techniques enhance the robustness and safety of LiDAR perception, enabling more reliable and trustworthy navigation in real-world settings. I will then introduce our recent progress on part-level segmentation in an open-vocabulary setting. Unlike object-level recognition, part segmentation presents additional challenges due to the ambiguity of defining what constitutes a part and the scarcity of annotated data. Yet, part-level understanding is essential for enabling robotic hands to interact meaningfully with objects in the real world. Our work addresses these issues with scalable methods that generalize beyond fixed taxonomies, allowing robots to flexibly interpret and manipulate novel objects. Together, these advances represent key steps toward resilient robot navigation and the development of versatile robots capable of seamless physical interaction with their environment.

 

 

[Biography]

Hyunjung Shim is an Associate Professor at the Kim Jaechul Graduate School of AI, KAIST, where she leads the Computer Vision and Machine Learning Laboratory (CVML Lab). She received her M.S. and Ph.D. in Electrical and Computer Engineering from Carnegie Mellon University, and her B.S. from Yonsei University. Before joining academia, she worked as a Research Scientist at Samsung Advanced Institute of Technology (SAIT). Her research focuses on vision-language foundation models, 3D vision,

trustworthy generative AI, and autonomous driving and robot perception. Dr. Shim serves as an Associate Editor for IEEE TPAMI and Area Chair for major AI conferences. She also serves as an Independent Director at Dongwon Industries Group.