Robot learning · SeoulTech

DongGeon (Lloyd) Kwak

I'm a fourth-year undergraduate in Computer Science and Engineering (minor in Applied AI) at Seoul National University of Science and Technology.

I work on robot learning. My current project corrects a pretrained diffusion policy on a robot whose joint has failed, without retraining. Before that, I worked as a software developer at two companies, and I'm a co-inventor on a Korean patent for crack detection with mobile robots.

DongGeon Kwak

Research

I'm interested in generalist robot learning: policies that transfer across robot bodies (cross-embodiment learning), keep improving after deployment by detecting and recovering from failures (lifelong robot learning), and carry out long-horizon manipulation with task and motion planning.

FAVOR rollouts
FAVOR: Fault-Aware Visuomotor Policy Adaptation via Ordered Retargeting
DongGeon Kwak
Manuscript in preparation, 2026

Training-free correction of a pretrained diffusion policy when a robot joint locks or loses range. Across 112 simulated fault conditions, restoring the end-effector position first raised mean success from 0.37 to 0.60 (0.11 to 0.48 with a locked joint). Real-robot tests on an xArm7 are in progress.

MTREA pipeline
MTREA: Multi-Task Retentive Embodiment Adaptation
Two-person team · DongGeon Kwak, contributor
Capstone Design, SeoulTech, 2026 (ongoing)

Does fine-tuning the π0.5 VLA on a new low-cost arm (SO-101) degrade its performance on held-out tasks? I co-designed the protocol and built teleoperation, remote training and policy serving; the evaluation is next.

Selected projects

From learning systems to software people can use.

Photo to brick model
MakeYourBrick: One Photo to a Buildable LEGO Model
SeoulTech ST startup club (team PixBricks), 2026

A web app that reconstructs an object from one photo with SAM2 and SAM 3D Objects and converts it into a buildable LEGO model (10,263 bricks for the queen shown). I worked on color mapping and shadow artifacts.

Patent

Real-time crack detection system for construction sites
K. Lee, G. Lee, D. Kwak
Korean Patent 10-2647135, registered March 2024
Mobile robots record video at construction sites, and cracks are detected and tracked across frames.

Experience & education

Research, industry experience and education
2025 – nowRobot Learning Research (FAVOR). Research feedback from Prof. Ji-Hyeong Han and participation in weekly lab meetings.
2022 – 2023Software Developer, Laver and CHASER
2022 – 2027B.E. Computer Science and Engineering, SeoulTech, minor in Applied AI; GPA 4.13/4.5; expected Aug. 2027

Get in touch

For research and engineering conversations, reach me at kwaklloyd@gmail.com.