Jongwook Jeon

I'm an undergraduate student in Yonsei University majoring in Electrical and Electronic Engineering. Currently, I am advised by Prof. Jongmin Lee.

My research interests center on building generalizable generative policies that can adapt across offline datasets, online interaction, and long-horizon real-world tasks. In particular, I focus on two directions:

  • RL Fine-tuning for Generative Policy Models:
    improving imitation-learned policies through RL fine-tuning, with a focus on flow-based and diffusion-based policy models.
  • Skill-based and Hierarchical Reinforcement Learning:
    composing reusable skills for long-horizon tasks, especially in VLA models and flow-/diffusion-based policies.

Email | CV | Github

profile photo

Publications

(*: indicates equal contribution)

OptiFlow
OptiFlow

We introduce OptiFlow, a novel approach for learning multimodal one-step flow policies using value-weighted optimal transport, enabling more effective and efficient policy learning in offline Reinforcement Learning.

MIMDICE
MIMDICE

We introduce MIMDICE, an off-policy algorithm that directly optimizes the mutual information with respect to the policy's stationary distribution via density ratio correction, enabling the use of arbitrary unlabeled data without requiring skill annotations.

Projects

These include side projects and unpublished research work.

AI Pharmacy Copilot for Robotic Medication Sorting xyz AI x Robotics Hackathon (Grand Prize)

Motivated by the repetitive workload and fatigue pharmacists often face in sorting and organizing medications, we developed an end-to-end AI robotic system that combines speech understanding, RGB-D perception, segmentation, VLM-based planning, and ROS2 control with safety and recovery mechanisms.

safs_small
safs_small

By modeling the optimizer and learning rate scheduler updates as a sequential decision process, we optimize the strategy via a combination of BC and PPO.

safs_small
FinPT: Finance Chatbot Yonsei 2nd GenAI (SilverPrize)
safs_small

Motivated by the critical issue of financial illiteracy among college students, FinPT is a web service designed to make financial knowledge accessible and user-friendly. By leveraging RAG technology, FinPT retrieves real-time data to provide students with up-to-date and easy-to-understand financial information.

Seminars

RL seminar
RL seminar

Delivered a comprehensive 5-week seminar series on Reinforcement Learning to the club team members in in Yonsei Data Science Lab Team. This series provided a theoretical explanation of Reinforcement based on CS285 & CAS4160.


Website source code credit to Dr. Jon Barron.