Dohyeok Lee

I'm a first year Ph.D. student at Cognitive Machine Learning Laboratory, advised by Prof. Jungwoo Lee, in Department of Electrical and Computer Engineering at Seoul National University (SNU). Previously, I received my M.S. in ECE from SNU in 2024 and B.S. in EE from KAIST in 2020.

Research Keywords: Robot Learning, Robotics, Learning from Demonstration (LfD), Reinforcement Learning (RL).

Email  /  CV  /  Google Scholar  /  Github  /  LinkedIn

profile photo

Research Vision

How can we leverage physics understanding to make robot learning truly generalizable?

My research focuses on empowering robots with physical understanding to achieve robust generalization across diverse manipulation scenarios — when agents truly understand the physics and affordances of their environment, they can adapt to novel scenarios beyond their training distribution. I develop algorithms that leverage spatiotemporal dynamics, combining novel view synthesis and dynamics prediction to enable robots to make informed decisions on unseen data. My vision is to create robotic systems that autonomously discover and apply physical principles to solve any manipulation task, achieving human-level adaptability in unstructured environments.

Publications

PontTuset SPQR: Controlling Q-ensemble Independence with Spiked Random Model for Reinforcement Learning
Dohyeok Lee, Seungyub Han, Taehyun Cho, Jungwoo Lee
NeurIPS, 2023
arXiv / code

We proposed SPQR, the first theoretically-grounded independence regularization for ensemble Q-learning based on random matrix theory. Our approach demonstrates significant improvements in reducing overestimation bias while achieving better computational efficiency in ensemble Q-learning.

PontTuset ARTificial Expressions: Human-Robot Interactive Drawing
Yejin Kim, Dohyeok Lee
CVPR Demo, 2023   (Best Demo)
code

We created ARTE, an interactive drawing system enabling collaboration between humans and robots. The system features a CLIP-based reward mechanism for real-time drawing state assessment and a reinforcement learning policy trained on diverse datasets using brush stroke simulation.

PontTuset Dynamics-Aligned Flow Matching Policy for Robot Learning
Dohyeok Lee, Jung Min Lee, Munkyung Kim, Seokhun Ju, Seungyub Han, Jin Woo Koo, Jungwoo Lee
CVPR Embodied AI Workshop, 2025
paper

We proposed Dynamics-Aligned Flow Matching Policy to address generalization to out-of-distribution scenarios by integrating random trajectory data (zero annotation cost) with expert demonstrations using dynamics model. Our iterative flow generation enables dynamics and policy models to mutually correct each other during training, combining the robustness of dynamics model with the flexibility of learning.

PontTuset View-Imagination: Enhancing Visuomotor Control with Adaptive View Synthesis
Dohyeok Lee, Munkyung Kim, Jung Min Lee, Seungyub Han, Jungwoo Lee
CVPR Embodied AI Workshop, 2025
paper

We proposed View-Imagination, a novel framework that dynamically selects optimal camera viewpoints for robotic manipulation using adaptive view synthesis. Based on our key insight that the most informative viewpoint is scene-dependent, View-Imagination trains a learnable viewpoint policy enabling robots to actively resolve visual ambiguities like occlusions.

Control of Furuta Pendulum with Reinforcement Learning
Dohyeok Lee, Usama Mohammad, Dong Eui Chang
ICCAS, 2019
video

Implemented a robust control system for the Furuta pendulum combining swing-up and balancing tasks using DDPG and PPO algorithms, with sim2real transfer and in-the-wild training.

Work Experience

Prior to my PhD, I gained hands-on robotics experience at two startups, building complete robotic systems from hardware to deployment. This practical background informs my current research on making learning algorithms work reliably in real-world settings.

Robotics Engineer
D.Hive (start-up), Daejeon, Korea
2020.10 - 2021.04

  • Built autonomous outdoor delivery robot from scratch, integrating LiDAR-camera fusion for robust navigation in urban environments
  • Led 10-engineer team across hardware/software, achieving successful deployment in real-world delivery scenarios

Robotics Engineer Intern
Crazing Lab (start-up), Pangyo, Korea
2019.06 - 2019.08

  • Built autonomous filming robot: hardware(frame,battery system), BLDC motor control system, and UART communication system
  • Implemented ROS system for motor control, IMU, LiDAR, and depth camera data processing

Open Source Contribution

PontTuset IMPALA
Open Source Contribution, 2024
code

Implemented IMPALA(Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures) in distributed system with ray, redis, and UDP

Nonlinear Controller (★20)
Open Source Contribution, 2021
code

Implemented nonlinear control (robust, adaptive, sliding mode) algorithms on two-arm manipulator simulator

PontTuset EKF (★14)
Open Source Contribution, 2021
code

Implemented EKF(Extended Kalman Filter) for sensor fusion of GPS and IMU data with Kitti dataset

RRT
Open Source Contribution, 2021
code

Implemented RRT(Rapid Random Tree) algorithms

Robotics Project

PontTuset Mobile Humanoid
Course Project at SNU (Actuation and Sensing Mechanisms for Robots), 2024
Collaborator: Dohyeok Lee, and 23 students

Developed wheel-based humanoid for navigation and object manipulation

Robot-AR system
In collaboration with Zer01ne(Hyundai Motor Company), 2021
Collaborator: Minyoung Kim, Yejin Kim, Dohyeok Lee, Junyoung Kim, Sunho Chang
video1 / video2

Developed Unity-ROS pipeline for AR visualization of Boston Dynamics Spot, enabling intuitive robot control system

PontTuset Vender
In collaboration with Art Center Nabi, 2020
Collaborator: Minyoung Kim, Dohyeok Lee, Seonguk Seo, Taewon Kang, Dahye Lee, Daeun Kim
video

Created A.I media artwork with A.I based emotion recognition and autonomous vending machine system

Autonomous Mobile Robot
In collaboration with robotics club MR, KAIST, 2018
Collaborator: Dohyeok Lee, Inyub Kim, Yongmin Lee, Dokyun Lee
video

Developed autonomous mobile robot with YOLO, Tmap API, GPS and compass sensor, etc.

PontTuset Hand-shape Manipulator with Teleoperation
In collaboration with robotics club MR, KAIST, 2017
Collaborator: Dohyeok Lee, Jaemin Cho, Jinsub Lee, Kiheon Sung

Developed hand-shape manipulator and glove-shape interface for teleoperation

PontTuset Marker-based Mobile Robot
In collaboration with robotics club MR, KAIST, 2016
Collaborator: Duckyu Choi, Hwijoon Lim, Dohyeok Lee

Developed mobile robot for marker-based localization and mapping

Other Research Project

Minimum distortion embedding for RL
Taehyun Cho, Dohyeok Lee, Jungwoo Lee
KICS Winter Conference, 2023
preprint / code

Proposed isometric regularization for RL to minimize distortion of latent space embedding

PontTuset Separated batch ensemble DQN
Dohyeok Lee, Jungwoo Lee
KICS Winter Conference, 2023
code

Proposed separated batch ensemble DQN for diversification of ensemble using separated batch for Bellman Q-target

PontTuset Genetic Algorithm for Surface Decomposition
Research Project, 2022
code

Implemented genetic algorithm for earth surface decomposition with arbitrary basis function

Simulator and Reinforcement Learning Algorithms for Surveillance/Reconnaissance
Changsik Lee, Dohyeok Lee, Dong Eui Chang
KIMST Conference, 2020

Developed simulation environment for surveillance/reconnaissance and reinforcement learning algorithms for surveillance agent

PontTuset 3D Box Fitting
In collaboration with RCV KAIST, 2018
Collaborator: Dohyeok Lee, Jaekook Hyun

Developed 3D box fitting algorithm for given point cloud data, collaboration with Hubo lab

Miscellanea

Service: Reviewer for NeurIPS 2025, CoRL Workshop 2024, ITW 2024

template adapted from here.