Robotics · Robot Learning · Manipulation
Robotics Engineering undergraduate at Hanyang University ERICA and undergraduate research intern at KITECH since July 2026. Current work focuses on teleoperation-based manipulation demonstrations, imitation learning, and data-efficient demonstration collection. Interested in combining data-driven robot-learning policies with robot kinematics and structured control.
Built a teleoperation pipeline from a physical OMY-L100 leader to a Franka FR3 in MuJoCo, including Cartesian pose retargeting, 6D-to-7DoF damped least-squares inverse kinematics, and null-space control. The platform is used to collect manipulation demonstrations, convert them to LeRobot format, and train and evaluate ACT and Diffusion Policy. Currently extending the platform with contact-force-based haptic feedback for manipulation experiments.
Studying ACT and Diffusion Policy in LeRobot-based manipulation tasks, including Push-T analysis and FR3 MuJoCo experiments using teleoperation demonstrations. Current work investigates the relationship between execution strategy, trajectory coverage, demonstration distribution, and policy behavior.
The Push-T study includes exploratory 50-episode comparisons of ACT execution configurations. FR3 results are reported separately as fixed-initial-condition pilot evaluations, not as a generalization benchmark.
My broader research interests include robot manipulation, teleoperation, learning from human demonstrations, imitation learning, and data-efficient robot learning. I am particularly interested in combining learning-based policies with kinematics- and control-based execution for robot manipulation.
navigation_stack_lab — Real-hardware ROS 2 navigation and system integration with SLAM, EKF sensor fusion, Nav2, DWB/MPPI experiments, and person-following logic. MPPI computation and parameters were optimized to improve the control loop from approximately 5–6 Hz to a stable 20 Hz.
rgbd-object-pose-estimation — RGB-D manipulation perception using an Intel RealSense D455, YOLOv8-seg, CAD meshes, and FoundationPose. The verified mean mask mAP50 is 0.929 across three industrial-part classes; this is a segmentation metric, not 6D-pose accuracy.
- 1st Place — 2026 Creative Comprehensive Design Competition
- Bronze Award — College of Engineering Capstone Design Competition
- Final — KAIST Mobility Challenge 2025
- Undergraduate Paper Award — ICROS 2026
- “Performance Comparison of DWB and MPPI Local Planners in ROS 2 Nav2 Environment” — ICROS 2026, poster, first author
- “Multimodal Perception and Monitoring System for Indoor Security Patrol Robots” — ICROS 2026, poster, Undergraduate Paper Award
ROS 2 · MuJoCo · LeRobot · ACT · Diffusion Policy · Python · C/C++ · PyTorch · Nav2 · SLAM · YOLOv8 · FoundationPose