Selected work in robotics, world models, and generative-image detection.
Bimanual manipulation
Connects learned motion plans to screw-decomposed impedance control for bimanual articulated-object manipulation. An RL policy uses object screw axes and wrench feedback to adapt compliance and reduce opposing forces between the arms.
World models and planning
Predicts dynamics and vision-language embeddings in a recurrent world model for MPC with text or image goals. Delta-score rewards support sequential subtask planning without training a separate policy for each goal.
Computer vision · Aegis AI
Led development of Aegis AI's generative-image detection system for the K-Startup Challenge.
Sim-to-real reinforcement learning
A self-balancing suitcase that shifts its center of mass under PPO control, trained in Isaac Lab and deployed on hardware. My contribution covered the simulation environment, policy training, and sim-to-real bridge.