I am a Ph.D. student at Data and Visual Analytics Lab (DAVIAN-LAB), advised by Prof. Jaegul Choo at Korea Advanced Institute of Science and Technology (KAIST).
Currently, I am a Deep Learning research intern at KRAFTON, mentored by Kangwook Lee. During my internship, I am developing a Co-playable Character (CPC), a unique type of in-game character that differs from traditional rule-based Non-Player Character (NPC). I am leading the development of the core algorithms for the CPC, which will be integrated into PUBG Battlegrounds.
My primary research objective is to develop a generalist agent capable of understanding its environment, making decisions, and autonomously learning from its experiences. I am currently addressing this challenge by integrating high-level reasoning using foundational models like Vision-Language Models (VLMs) or Large Language Models (LLMs) with low-level control through Reinforcement Learning (RL) or Behavior Cloning (BC).
Email / CV / Google Scholar / Linkedin / Github
Designed network architectures that steer convergence toward simple functions which allows to scale up parameters in RL.
We present DoDont, an instruction-based skill discovery algorithm designed to combine human intention with unsupervised skill discovery. DoDont learns diverse behvaiors while following the behaviors in "do" videos while avoiding the behaviors in "don't" videos.
To allow the network to continually adapt and generalize, we introduce Hare and Tortoise architecture, inspired by the complementary learning system of the human brain.
We introduce CoIn, a lightweight and simple add-on module, which effectively adapts pretrained Vision Transformers for visuo-motor control.
We introduce DISCO-DANCE, a skill discovery algorithm focused on learning diverse, task-agnostic behaviors. DISCO-DANCE addresses the common limitation of exploration in skill discovery algorithms through explicit guidance.
For sample-efficient RL, the agent needs to quickly adapt to various inputs (input plasticity) and outputs (label plasticity). We present PLASTIC, which maintains both input and label plasticity by identifying smooth local minima and preserving gradient flow.
We gathered data from 280,000 matches played by the top 0.3% rank players in Korea for League of Legends. From this, we developed DraftRec, a personalized champion recommendation system aimed at maximizing players' win rates.
Template based on Hojoon's website.