Sample-Efficient Diffusion-based Reinforcement Learning with Critic Guidance

Researchers introduced CGPO, a training-free critic-guided diffusion policy optimization method for reinforcement learning.

CGPO steers action generation toward high-value regions and uses those actions as regression targets.

It achieved state-of-the-art results on five MuJoCo locomotion benchmarks and strong performance on Franka robot grasping tasks.

The method improves sample efficiency by better balancing exploration and exploitation in diffusion-based RL.