- (2026.07 ~) Korea University, Seoul (Department of Artificial Intelligence)
- (2020.03 ~ 2026.07) Dongguk University, Seoul (B.S. in Department of Computer Science and AI)
- Surgical video perception. I study how to detect and segment regions marked by indocyanine green (ICG) in endoscopic video and track the same tissue through deformation, occlusion, and changes in fluorescence.
- Data-efficient robot learning. I’m interested in what first-person videos of human manipulation can teach a robot, especially when only a small set of robot demonstrations is available.
- Motion planning and control. I use MuJoCo to study model predictive path integral control (MPPI) for surgical manipulators. One constraint I work with is that the instrument shaft must pivot around a fixed entry point as its tip moves.
- Medical vision–language models. In TaSA(Target-State Matching Preference Optimization for Vision--Language Alignment in Context-Sharing Domains), I helped construct pairs of medical images with and without a target lesion and tested whether image–text similarity changed when the lesion was removed.
- Tracking tissue after fluorescence fades. I compared point tracking with SAM 2 mask propagation on endoscopic video. In one low-contrast sequence, the point tracker lost all 16 tracked points. The mask persisted longer but sometimes spread onto neighboring tissue. That distinction matters for a surgical overlay: a mask can stay visible even after it has lost the original target.
- Needle reaching in simulation. I built a teleoperation demo for the da Vinci Research Kit’s patient-side manipulator. The demo displays two errors as the instrument moves: how far the tip is from the needle target and how far the shaft deviates from its intended pivot. Explore the MuJoCo demo
- Predicting surgical video from a single frame. I compared DreamDojo-2B and Cosmos-Predict 2.5 on the same surgical clips, giving each model the first frame and asking it to predict the next 12. Using the same training budget let me examine what a model pretrained on first-person human video could transfer to a surgical scene.
⭐️ I write about my experiments and the papers behind them on my research blog. ⭐️
- Robot Intelligence: physics-based simulation, trajectory optimization, and motion learning.
- Human-Robot Alignment: imitation learning, VLA models, and embodiment-agnostic representations.
- Medical AI & Computer Vision: medical vision-language models and data-efficient segmentation.
- Analogical Reasoning: cross-domain methods inspired by robotics, physics, and nature.


