Dohyeok Lee

I'm a Ph.D. student in Electrical and Computer Engineering at Seoul National University (SNU), advised by Prof. Jungwoo Lee. Currently, I am a visiting Ph.D. student working with Prof. Tapomayukh Bhattacharjee at Cornell University.

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, Contact-Rich Manipulation, Visuomotor Control, Dynamics Modeling, Reinforcement Learning

Email  /  CV  /  Google Scholar  /  GitHub  /  LinkedIn

Research Vision

Can physical understanding become the foundation for generalizable robotic dexterity?

My research is driven by a simple belief: robots should not merely fit correlations in data, but learn the physical structure that governs interaction. Unlike language, physical interaction is governed by laws of nature, and these laws can provide a foundation for generalization beyond the training distribution.

Two questions guide my work: how to represent physics, and how to move beyond prediction toward understanding. I study these questions through dynamics modeling, novel view synthesis, and contact-aware control, with the goal of building robots that internalize causal physical structure and achieve dexterous manipulation in unseen, occluded, and contact-rich settings.

Selected Publications

Learning latent actions from unlabeled videos is often confounded by viewpoint-dependent visual changes. MVP-LAM uses cross-viewpoint reconstruction on synchronized multi-view videos to learn action-centric latent actions, improving VLA pretraining and downstream manipulation on SIMPLER and LIBERO-Long.

Whole-arm manipulation is challenging because contact states and interaction forces are partially observed and tightly coupled with robot motion. TACTIC combines RGB-D, proximity, and distributed tactile sensing with contact-aware MPC and hybrid dynamics, enabling safer contact-rich manipulation on simulated and real-world whole-arm tasks.

We propose DistLCB, an algorithm for multi-armed bandits under heavy-tailed rewards that achieves asymptotic optimality for Lipschitz risk measures and extends to Pareto-optimal multi-risk objectives.

We proposed SPQR, the first theoretically-grounded independence regularization for ensemble Q-learning based on random matrix theory. By ensuring proper independence, our approach significantly improves generalization to out-of-distribution data across diverse RL tasks.

Demos and Workshops

We created ARTE, a real-world human-robot collaborative drawing system with multi-modal goal-conditioned policy. This work validates our generalization approaches in actual deployment, demonstrating robust performance across diverse drawing scenarios and human interaction patterns.

We address temporal understanding by learning explicit dynamics and mutually correct dynamics models and policy by iterative flow generation. Our key insight: random trajectory data (zero annotation cost) enables robust generalization beyond expert demonstrations to out-of-distribution scenarios.

View-Imagination addresses visual ambiguity in fixed-camera robot manipulation by synthesizing and selecting task-informative viewpoints, improving robustness to occlusions and unseen spatial configurations.

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

Industry Experience

Hands-on robotics engineering experience at early-stage robotics startups, building real-world systems from hardware integration to deployment.

template adapted from here.