Robotics · Humanoid · VLA
M.S. student at Sungkyunkwan University, working with the Robotics & Intelligent Systems Lab, advised by Prof. Hyungpil Moon.
I study how robots can learn useful behavior from vision and demonstration. My current interests are vision-language-action models, viewpoint-robust imitation learning, and whole-body humanoid teleoperation.
How can a robot policy stay reliable when the camera moves, the object is occluded, or the embodiment changes?
My work sits between robot learning and systems engineering. I am particularly interested in policies that understand where they are looking, actively seek better observations, and transfer what they learn to real humanoid platforms.
Small updates from ongoing work.
Integrating hand and arm control into a humanoid VR teleoperation pipeline.
Studied visual subgoal generation and hierarchical manipulation policies.
Investigated camera conditioning for view-invariant imitation learning.
Reviewed active perception methods for VLA-based robotic manipulation.
Implemented VR teleoperation for an Agibot humanoid platform and developed an automated tensile-testing workflow with a π0.5 policy.
My contribution Agibot VR teleoperation integration and π0.5 task development for tensile-test automation.
Developing a K-logistics humanoid robot that combines a high-sensitivity robotic hand with multimodal AI foundation models. The work deploys imitation-learning and vision-language-action policies for robot manipulation.
My contribution Built an Isaac Sim environment and task assets that reproduce the real workcell, enabling demonstration-data augmentation for policy training.
An end-to-end study of dataset preparation, fine-tuning, inference, and robot-side evaluation. The goal is to identify which assumptions fail when a generalist VLA meets a custom robot setup.
Focus Data representation, camera setup, and deployment failure analysis.
Exploring robot policies that remain reliable as viewpoint and occlusion change. This work connects camera conditioning, active perception, and task-aware view selection.
Research question Should the policy become view-invariant, or should the robot learn where to look?
Published in arXiv preprint arXiv:2607.20912, 2026
Sungkyunkwan University
Robotics & Intelligent Systems Lab
Major GPA 4.33 / 4.5
Major GPA 4.33 / 4.5
Vision Group, Robot Dynamics & Motion Control Team
Developed box segmentation and dimension-estimation components for depalletizing unknown box configurations.
Implemented an image-data collection pipeline with automated data gathering and auto-labeling workflows.
Built data-gathering and visualization workflows, and evaluated manipulation in simulation using VR teleoperation.
Doosan Robotics · Excellent internship evaluation
ROKEY Bootcamp · Recognized for team performance and leadership
Graduated top of the class
Ministry of Science and ICT
PyTorch, VLA, imitation learning, computer vision
ROS 2, teleoperation, manipulation, humanoid control
Python, Linux, Git, SolidWorks, robot data pipelines
Interested in robot learning, VLA systems, or humanoid?
yosmon@g.skku.eud ↗