About me
Hi, I’m Co Yong, a PhD student in the Data Science Degree Program at National Taiwan University and Academia Sinica under the supervision of Prof. Shao-Hua Sun. My research focuses on offline RL, offline imitation learning and model-based planning. I’m recently exploring zero-shot RL.
Publications
Less Tuning, Better Planning: Simplifying Offline Model-Based Planning
Published in ICML 2026 workshop on Decision-Making from Offline Datasets to Online Adaptation: Black-Box Optimization to Reinforcement Learning, 2026
Offline model-based planning can adapt policies at test time, but performance depends on a planning horizon and action proposer that are often tuned online. We propose SHARP (Soft Horizon AggRegation for Planning), which weights multi-horizon returns by ensemble dynamics uncertainty to avoid fixed-horizon tuning. SHARP-BC pairs this with a simple behavior-cloning action proposer, matching or beating baselines with less hyperparameter search.
Restoring Noisy Demonstration for Imitation Learning With Diffusion Models
Published in IEEE Transactions on Neural Networks and Learning Systems, 2026
Most imitation learning methods assume perfect expert demonstrations, yet real data is often noisy. We propose a filter-and-restore framework that isolates clean samples and uses conditional diffusion models to recover noisy demonstrations. Experiments on robot arm manipulation, dexterous manipulation, and locomotion show consistent improvements over existing methods, with ablations confirming robustness to diverse noise types and levels.