Stony Brook University Computer Science

Prof. Ukcheol Shin will be giving a talk on "Robust Physical AI in a Challenging Environment."

Please find the seminar details below:
Title: Robust Physical AI in a Challenging Environment
Date & Time: Friday, May 22, 2 PM
Venue: B204

Bio
Prof. Ukcheol Shin is an Assistant Professor at the School of Energy Engineering and Institute for Energy AI at KENTECH.
Before joining KENTECH, he was a postdoctoral researcher at Bot Intelligence Group, Robotics Institute, CMU with Prof. Jean Oh. He obtained my Ph.D. and M.S. at KAIST, advised by Professor In So Kweon. He is a recipient of the Best Student Paper Award from WACV 2023, the 29th Samsung Humantech Paper Award, the 1st Place Award at 2024 CVPR WAD workshop, and the Best Poster Award at 2025 ICRA TIRO workshop. Also, he co-organized the “Thermal Infrared in Robotics” workshop at ICRA 2025 and the “Multi-spectral Imaging for Robotics and Automation” workshop at ICCV 2025.
Prof. Ukcheol Shin's research focuses on developing a robust physical AI that can perceive, understand, and navigate the dynamic world in challenging conditions, with a specific interest in spatial/semantic perception in extreme conditions, self-supervised learning, deep reinforcement learning, multi-sensor fusion, and vision-language navigation/manipulation. 

Abstract
Building robust Physical AI systems that operate reliably in real-world, adverse environments remains a fundamental challenge in robotics and embodied intelligence. This talk presents our recent efforts toward developing resilient visual and control intelligence under challenging conditions such as low visibility, smoke, and extreme lighting.
First, we explore thermal imaging as an alternative and complementary sensing modality for robust visual perception. We discuss self-supervised learning approaches tailored for thermal data to enable geometry- and semantics-aware representations without heavy annotation. Next, we introduce reinforcement learning strategies designed to enhance control robustness under edge cases and distribution shifts, enabling safer and more adaptive behaviors in unstructured environments.
Finally, we outline future research directions toward robust visual locomotion and language-aware embodied intelligence, aiming to bridge perception, decision-making, and action for real-world Physical AI systems operating in the wild.