DexDICE: Reconciling Safety and Performance via Dual-Expert Offline Imitation Learning NEW
Seokin Seo์์์ธ
Postdoctoral Researcher
Advanced Robot Control Lab, Center for Humanoid Research, Korea Institute of Science and Technology (KIST)
I am a postdoctoral researcher at KIST working with Dr. Yisoo Lee (since March 2026). I received my Ph.D. from KAIST AI, co-advised by Prof. Kee-Eung Kim and Prof. Hongseok Yang. My research focuses on learning robust decision-making policies from data, with applications to humanoid robot learning and control.
Education
-
2026Ph.D., Kim Jaechul Graduate School of AI, KAIST (Advisors: Kee-Eung Kim, Hongseok Yang)DissertationRobust Imitation Learning under Data Collection Biases
-
2019M.S., School of Computing, KAIST (Advisor: Kee-Eung Kim)ThesisA Study on Generating Explanations for Reinforcement Learning Policies in Control Tasks
-
2017B.S., Computer Science & Mathematics, Sogang University
Research Interests
Reinforcement LearningImitation LearningPhysical AI
My research lies at the intersection of reinforcement learning and imitation learning, with the broad goal of building sequential decision-making agents that remain reliable when the data they learn from is imperfect. During my Ph.D., I worked on offline imitation learning that is robust to covariate shift and data collection biases (DrilDICE, NeurIPS’24; PALR, NeurIPS’23; DemoDICE, ICLR’22), and more recently on learning safe behavior without hand-specified cost functions (DexDICE, NeurIPS’26).
At KIST, I am extending this line of work toward physical AI, where these methods meet real robots such as humanoids, manipulators, and mobile robots.
Publications
* equal contribution, † equal advising.
International Conferences & Preprints
2024
2023
Domestic Conferences & Journals
Awards
- Gold Reviewer Award, ICMLJul. 2026
- Outstanding Paper Award, Korea Software Congress (KSC)Feb. 2023
Academic Talks
- Robust and Safe Offline Imitation Learning via Distribution Matching
Locomotion and Manipulation Technology Workshop, Korea University2026 - Mitigating Covariate Shift in Behavioral Cloning via Robust Stationary Distribution Correction
AI Tutorial Series, KAIST2025
Academic Service
- Reviewer, NeurIPS2021, 2024, 2026
- Reviewer, ICML2025, 2026 Gold Reviewer
- Reviewer, ICLR2025, 2026
- Reviewer, ICRA2027
Teaching Experience
- Guest Lecturer, Mathematics for AI (AI503), KAIST
Lecture: Distribution Shift in Imitation LearningFall 2025 - TA, Introduction to Reinforcement Learning, KAIST Machine Learning Engineer Bootcamp2022
- TA, Mathematics for AI (AI503), KAISTFall 2021
- TA, Dialogue System, KAISTโSamsung DS AI Expert Program2020
- TA, Machine Learning (CS376), KAISTFall 2019
- TA, Data Structure (CS206), KAISTSpring/Fall 2018