Research Project
Reinforcement Learning (Rl) Methods For Care-Aimed Assistive Robots
This project develops reinforcement learning (RL) methods for assistive robots in home-care, hospital, and care-center environments. It first focuses on zero-shot RL generalization for robotic manipulation, aiming to enable robot arms to pick, place, and deliver diverse unseen objects such as medications, IV drips, injections, and daily necessities without task-specific retraining. To make this feasible in real-world care settings, the project then studies cross-domain RL, transferring knowledge from data-rich, low-cost source domains such as simulation to data-scarce, high-cost target domains such as hospital or home-care robots. This reduces the burden and risk of collecting large-scale real-world robot data. Finally, the project incorporates human-like RL so that robot action sequences better align with human behavior. By distilling demonstrations into reusable macro actions and optimizing human-like trajectories, the robots can behave more naturally and support seamless human–robot collaboration in sensitive care environments.