Donghu Kim

logo_uni_tue          logo_uni_tue

Hi there, nice to meet you!

My name is Donghu Kim. I am on a Master's Degree program in KAIST (advised by Jaegul Choo), studying reinforcement learning and embodied AI.

Towards building a generalist robotic agent, I am a huge supporter of the belief that RL will be the one that generates the low-level data [1, 2]. In this direction, I am invested in pushing the absolute limit of efficiency in RL for control: Can we make RL work with only 1K samples? Can we do it within an hour? As far fetched as the goal may seem, there are so many exciting components we can tackle, including feature learning, exploration, architecture design, optimizer, task transfer, etc.

I still have a long long way to go; if you want to discuss anything research related, I'd be more than happy to be engaged!

Email  /  CV  /  Google Scholar  /  Github

profile photo

Work Experience

Krafton
Krafton AI June 2025 - Present
Research Intern
  • Physical Intelligence Team (TBD)

Publications

[ Selected / All ]
toxchat
Langauge Agent
Building Resource-Constrained Language Agents: A Korean Case Study on Chemical Toxicity Information
Hojun Cho, Donghu Kim, Soyoung Yang, Chan Lee, Hunjoo Lee, Jaegul Choo.
Under review.
arXiv

We build a light-weight LLM agent that answers chemical toxicity questions based on the Korean Tox-Info database.

simbav2
Reinforcement Learning
SimBaV2: Hyperspherical Normalization for Scalable Deep Reinforcement Learning
Hojoon Lee*, Youngdo Lee*, Takuma Seno, Donghu Kim, Peter Stone, Jaegul Choo.
NeurIPS'25, Spotlight.
arXiv / project page / github

We further regularize SimBa architecture by projecting both parameters and features onto a hypersphere, leading to better scaling properties in model size and compute.

simba
Reinforcement Learning
SimBa: Simplicity Bias for Scaling Up Parameters in Deep Reinforcement Learning
Hojoon Lee*, Dongyoon Hwang*, Donghu Kim, Hyunseung Kim, Jun Jet Tai, Kaushik Subramanian, Peter R. Wurman, Jaegul Choo, Peter Stone, Takuma Seno.
ICLR'25, Spotlight.
arXiv / project page / github

We propose a well-regularized architecture that avoids overfitting, allowing parameter and compute scale up in RL.

preprint2024dodont
Reinforcement Learning Skill Discovery
Do’s and Don’ts: Learning Desirable Skills with Instruction Videos
Hyunseung Kim, Byungkun Lee, Hojoon Lee, Dongyoon Hwang, Donghu Kim, Jaegul Choo
NeurIPS'24.
arXiv / project page

We present DoDont, a skill discovery algorithm that learns diverse behaviors while following the behaviors in "do" videos while avoiding the behaviors in "don't" videos.

icml2024atari-pb
Reinforcement Learning Pre-training
ATARI-PB: Investigating Pre-Training Objectives for Generalization in Pixel-Based RL
Donghu Kim*, Hojoon Lee*, Kyungmin Lee*, Dongyoon Hwang, Jaegul Choo.
ICML'24.
arXiv / project page / github / poster

We investigate which pre-training objectives are beneficial for in-distribution, near-out-of-distribution, and far-out-of-distribution generalization in visual reinforcement learning.

icml2024hnt
Reinforcement Learning Plasticity
Slow and Steady Wins the Race: Maintaining Plasticity with Hare and Tortoise Networks
Hojoon Lee, Hyeonseo Cho, Hyunseung Kim, Donghu Kim, Dugki Min, Jaegul Choo, Clare Lyle.
ICML'24.
arXiv / github

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.


Study Materials

Note: These slides are made for studying purposes only, and likely have got something wrong here and there. If you happen to find some, feel free to make fun of me via e-mail :).

Honors & Awards

Services (Reviewer)

Template based on Hojoon Lee's website.