Undergraduate Researcher @ CSI-Agent Lab, Sungkyunkwan University B.S. in Software Engineering · Double major in Electrical & Electronic Engineering
I work on robot learning and Vision-Language-Action (VLA) models — teaching robots to handle deformable, unstructured objects in the real world. Most of my time goes into bimanual manipulation pipelines, imitation learning, and figuring out why the policy that worked in sim doesn't work on hardware.
Currently interested in: generalization in VLA policies, co-training with multimodal data, and sim-to-real transfer for deformable object manipulation.
VLA Co-training Ablation · CSI-Agent Lab, 2026–
Investigating whether co-training a pretrained VLA policy on robotic demonstrations + VQA data
improves generalization to unseen settings. Four-condition ablation on SmolVLA over RoboTwin 2.0
tasks; I own the robot-data-only baseline and the Easy/Hard evaluation protocol.
SmolVLA RoboTwin 2.0 PyTorch
Bimanual Laundry Folding — KAIA Physical AI Challenge · Top award, final round · prism-sanhak
Bimanual SO-101 system for folding laundry end-to-end. Built the teleoperation data collection
system and the cube grasping task; contributed to the folding pipeline via DAgger-based
human-in-the-loop fine-tuning. Solved USB bandwidth contention across host controllers for stable
multi-camera/servo operation, and added a CLI recording fallback for teleop dropouts.
Integrated a ViT-based stage transition classifier into the high-level orchestrator.
LeRobot SO-101 DAgger ViT
A note on the team repository: the codebase was pushed in bulk by two members once the project was complete, so the commit history on GitHub doesn't reflect who wrote what. My contributions are as described above — happy to walk through any of it in detail.
RoboTwin 2.0 Reproduction · CSI-Agent Lab
Reproduced the RoboTwin 2.0 benchmark across all paper tasks with ACT and Diffusion Policy on
Aloha-AgileX. Easy 48% vs. paper 56%, Hard 0% vs. 3% — both within statistical error, independently
confirming the paper's domain randomization findings.
ACT Diffusion Policy Docker Vulkan
WishKit · 🥇 1st place / 14 teams, SKKU Software Studio · soft-studio-team
A Korean e-commerce wishlist app that treats your wishlist as a decision queue rather than a
bookmark folder. I own the price extraction and parsing engine — cross-mall entity resolution and
price tracking. Also shipped Kakao Share SDK integration and the Android build pipeline.
Flutter FastAPI Firebase
KOSPI200 Momentum Backtest
Full backtest pipeline for a 12-1 momentum factor on KOSPI200 — equal-weighted quintile portfolios
with monthly rebalancing. Next up: pairs trading with cointegration.
Python pandas pykrx
Languages Python · C++ · Dart · Verilog ML / Robotics PyTorch · LeRobot · ACT · SmolVLA · GR00T · Diffusion Policy · OpenCV Infra Docker · Linux · SSH/Remote-SSH · conda App Flutter · FastAPI · Firebase
Korean · English · Chinese