DiffSkill: Improving Reinforcement Learning through diffusion-based skill denoiser for robotic manipulation

Siao Liu, Yang Liu, Linqiang Hu, Ziqing Zhou, Yi Xie, Zhile Zhao, Wei Li, Zhongxue Gan

Abstract: Highlights•We propose a diffusion-based skill learning framework, named DiffSkill.•We present a Cycle Anneal Scheduler to dynamically tune denoising iterations.•We unveil an Online Momentum Smoothing strategy to avoid action oscillations.•DiffSkill exhibits remarkable performance improvements to previous methods.