CFG-Policy: Unified Diffusion Policy Learning Framework via Classifier-free Guidance

This paper presents CFG-Policy, a unified and plug-and-play framework for diffusion policy learning that separates exploration and exploitation into two generative policies. By combining their score or velocity fields through classifier-free guidance, the framework aims to improve task performance while preserving behavioral diversity. Results are shown on 5 MuJoCo tasks using QSM and QVPO, and on 4 RoboTwin manipulation tasks using Flow-SDE+PPO.