Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising

Haoyu Lei (The Chinese University of Hong Kong) · Yinchuan Li (Huawei Technologies Ltd.) · Yasser Khalil (Noah's Ark Lab, Huawei Technologies Ltd.) · Mateo Clémente (Huawei Technologies Ltd.) · Leo Brunswic (Huawei Technologies Ltd.) · Yang · Xuan Zhao (City University of Hong Kong) · Amir Rasouli (Huawei Technologies Canada)
action distributionsd4rldenoising processdiffusion modelsdiffusion policyembodied aiexpert demonstrationsgenetic denoisinginference budgetsneural function evaluationsout-of-distribution riskperformance stabilitypopulation-based samplingrobomimicrobotic manipulation

Diffusion models, such as diffusion policy, have achieved state-of-the-art results in robotic manipulation by imitating expert demonstrations. While diffusion models were originally developed for vision tasks like image and video generation, many of their inference strategies have been directly transferred to control domains without adaptation. In this work, we show that by tailoring the denoising process to the specific characteristics of embodied AI tasks—particularly the structured, low-dimensional nature of action distributions