Projects

• Proposed the initial GR00T vision-language-action (VLA) teacher–student design; the team moved to a faster modular pipeline when speed became the constraint.
• Trained the YOLO26 segmentation model and collected real D415 data; on held-out real viewpoints it found all 144 targets with no false positives, where the synthetic-only model had missed every guide.
• Before the competition, added the containerized SDK runtime, TCP position control, lens and yaw calibration, force-controlled gripping, shorter motion routes, and operator-console updates.
• Responsible for model training, code verification, and data collection.

Technologies: YOLO26 (instance segmentation) · Intel RealSense D415 · Rainbow Robotics RB10 · Zimmer GEH6060IL Gripper · MuJoCo · Python · Docker · NVIDIA GR00T (initial design)

• Launched on September 25, 2026 with 18 guided experiments across PID, lead/lag, LQR, Kalman filtering, and LQG, plus an 18-chapter bilingual Lecture Room.
• Runs a deterministic nonlinear simulation in a Web Worker and checks the motor model against an exact sampled linear-system reference computed with SciPy.
• Lets learners pin runs for comparison, save them locally, and share reproducible experiments as versioned URLs, JSON, or CSV.

Technologies: TypeScript · React · Next.js · Web Workers · ECharts · Cloudflare Workers

• Extracted frames from 720p video, estimated camera poses and sparse geometry with COLMAP, undistorted images, and trained 3DGS reconstructions.
• Compared reconstructions for a doll-on-table, table-only, and room scene; the best doll-on-table result reached PSNR 32.14.
• Showed that camera poses were essential to reconstruction quality, while sparse points served as initialization rather than ground-truth labels; random-point initialization reduced PSNR from 32.14 to 24.54.
• Identified COLMAP as the primary pipeline bottleneck: the 6 fps doll scene took about 51 minutes for COLMAP versus about 16 minutes for 30k 3DGS iterations.

Technologies: COLMAP · 3D Gaussian Splatting · CUDA · Conda · FFmpeg

• Collected and organized camera data directly, then annotated the dataset for model training.
• Trained a perception model from the labeled data and used the resulting pipeline to design an autonomous-driving algorithm.
• Learned how label balance, data coverage, and overall data quality can determine model behavior throughout the training process.
• Approached annotation as a model-aware task, including bounding boxes for small objects when they matter to the intended driving behavior.

Technologies: Computer Vision · Data Annotation · Object Detection · Python · OpenCV

• Formulated the maze as a Markov decision process and implemented Q-Learning and SARSA baselines, both of which reached the 15-step optimal path.
• Built a DQN with a three-layer CNN, experience replay, target network, RMSprop optimization, and epsilon-greedy exploration for 64 × 64 local RGB and grayscale observations.
• Showed that grayscale input retained the optimal 15-step path while training about 15% faster than RGB; the best tested configuration reached the target in 761 episodes with epsilon decay 0.95.
• Analyzed an RGB early-stopping model that appeared to converge by training reward but failed deterministic testing, highlighting the need for policy-level evaluation.

Technologies: Python · PyTorch · MiniGrid · Reinforcement Learning

• Awarded 2nd place at the Energy Up! 2026 New Energy Industry Idea Lab Competition, ranking 2nd among 13 teams.
• Received the President of Korea Electrical Engineers Association Award at the 2026 KIPE IE Competition, ranking 4th among 35 teams.
• Integrated transmitter and receiver modules into a ROBOTIS TurtleBot and redesigned its battery and electrical system.
• Proposed waypoint charging instead of a fully embedded track and designed the corresponding driving logic.

Technologies: WPT · Raspberry Pi · ROS · Arduino · Fusion 360 · Isaac Sim

• Designed for an 800 Vdc input, 400/500/600 V outputs, and 25–100% load conditions while maintaining zero-voltage switching across the specified range.
• Derived resonant-network parameters and voltage-gain curves, then estimated MOSFET and diode losses and converter efficiency for each load condition.
• Built a PSIM switching model and verified operating waveforms, ZVS, output ripple, load-dependent efficiency, and weighted efficiency.
• Compared a MATLAB parameter sweep with PSIM results, traced the discrepancy to harmonic content omitted by the analytical approximation, and selected the final parameters from waveform-based simulation.
• Verified dead time and adjusted resonant frequency and capacitor values under switching-frequency and commercially available component constraints.

Technologies: PSIM · MATLAB

• Implemented a PV array, boost converter, and P&O and Incremental Conductance MPPT control.
• Modeled a wind turbine, PMSM, rectifier, buck converter, and optimal-current control.
• Analyzed low-wind limitations and tested a buck-boost topology to improve tracking.

Technologies: MATLAB · Simulink

• Used the VisionProTeleop project with Apple Vision Pro and MuJoCo to support teleoperation-based expert-data generation.
• Generated expert demonstrations for the KETI Allegro dual-arm cashier task.
• Reviewed reward behavior during policy experiments to check alignment with the intended manipulation and grasping task.

Technologies: Apple Vision Pro · VisionProTeleop · MuJoCo · Imitation Learning · Allegro Hand

• Implemented explainable defect analysis with a vision LLM, providing the basis for each classification decision.
• Designed prompt-editing controls so users could iteratively improve performance without changing the application code.
• Combined OpenCV edge detection with a vision LLM to detect fine defects and contextual anomalies.
• Built a FastAPI/React interface with prompt editing, real-time monitoring, and F1 score visualization.
• Used AI-assisted development to iterate quickly on the FastAPI/React interface and application features within the hackathon timeframe.

Technologies: LangChain · OpenCV · FastAPI · React · GPT-4o-mini · Python

• Implemented real-time fault-injection scenarios in Simulink.
• Reduced convergence from 3,900 to 850 episodes, a 4.6× improvement in training efficiency.
• Submitted the work to the Simulink Student Challenge 2025 under the fault-tolerant control theme.

Technologies: MATLAB · Simulink · Reinforcement Learning Toolbox · Simscape Multibody

• Write structured learning notes on diffusion models and Diffusion Policy, from core concepts to their use in robot learning.
• Cover World Model research topics alongside reinforcement learning, imitation learning, and robotics papers.
• Document practical study and research context so theoretical material remains connected to experiments and implementation.

Technologies: Tistory · Robot Learning · Diffusion Policy · World Model

• Designed the world-model-conditioned Diffusion Policy architecture, experiment plan, and implementation while learning from a senior coauthor.
• Evaluated the policy on real Franka Emika Panda pick-and-place tasks, including a deformable object whose graspable regions change with orientation.
• Kept added inference latency to approximately 70 ms; presented the paper orally at ICCAS 2026 on Oct 28, 2026.

Technologies: Diffusion Policy · GRU World Model · FiLM Conditioning · Imitation Learning · Franka Emika Panda

• Designed a three-layer pipeline for wake-word-triggered command recognition, tray-level instrument and inventory recognition, and imitation-learning-based pick-and-place.
• Defined a closed-vocabulary voice workflow that activates STT only after the “Hey Mary” wake word and extracts an instrument name from the subsequent utterance.
• Selected placement in a designated delivery zone instead of direct robot-to-human handover to reduce contact paths and keep the manipulation task feasible.
• Planning the SO-ARM101 cart layout around its effective reach, with an overhead tray camera and a wrist camera for instrument recognition and manipulation.
• The manipulation-policy architecture remains under evaluation; this project does not claim readiness for real operating rooms, patients, or clinical instruments.

Technologies: SO-ARM101 · LeRobot · Whisper · VAD · Computer Vision · Imitation Learning

• EXIF Framer displays digital-camera capture metadata, including ISO, aperture, and shutter speed, in a photo frame.
• MEOW PDF Merger, first developed in December 2025, interleaves documents scanned with a single-sided feeder into a duplex-ordered PDF.

Technologies: TypeScript · Python · Vercel

• Built, deployed, and continue to operate Notion2Blogger with an AI-assisted development workflow; it converts Notion content for Blogger, including formulas and code blocks.

Technologies: Next.js · Python · TypeScript · Vercel · Supabase · Figma

• Designed the electrical power path around a 12 V, 30 A SMPS and DC-DC buck converters to supply the prototype subsystems.
• Built an ESP32-based mobile control page and used relay modules to switch the device functions from the web interface.
• Integrated prototype functions including PTC heating, ultrasonic humidification, UV lighting, and vibration, with fusing and supporting passive components in the electrical design.
• Coordinated rapid mechanical enclosure fabrication using acrylic panels, hinges, handles, and assembly materials during the three-day build.

Technologies: ESP32 · HTML/CSS/JavaScript · Relay Module · SMPS · DC-DC Converter

• Authored a Korean technical paper and modeled changes in crash scenarios with MATLAB.

Technologies: MATLAB

• Preprocessed acceleration, yaw-rate, GPS Speed2D, wheel-speed, and position data at 100 Hz; applied a high-pass Butterworth filter and converted between geographic and local ENU coordinates.
• Implemented and compared ten velocity-estimation approaches: four GPS differentiation methods, five IMU integration methods, and KF-RK4 sensor fusion.
• Found the 3-point central difference to be the most reliable GPS differentiation method, while the 5-point method was degraded by edge effects and noise sensitivity.
• Reduced IMU integration drift by fusing RK4 predictions with reference measurements through a Kalman Filter, achieving a velocity RMSE as low as 0.04 and the lowest path RMSE among the tested methods.
• Reconstructed the vehicle path using estimated velocity and integrated yaw rate, and analyzed late-trajectory divergence caused by accumulated heading drift; selected as an outstanding project team.

Technologies: MATLAB