Statistical Evaluation of Noise Schedulers for Diffusion Policy in Robotic Manipulation Tasks

This thesis investigates the impact of noise scheduler design in Denoising Diffusion Probabilis-tic Models (DDPMs) applied to robotic manipulation tasks. While DDPMs applied to robotics,namely diffusion policies, are recognized for modeling multimodal action distributions, thechoice of noise schedule in this field remains underexplored from a statistical standpoint. Wesystematically compare three commonly used schedulers (linear, cosine, and sigmoid) acrosstwo benchmark manipulation tasks (PickCube and StackCube) in the ManiSkill framework un-der low and high data regimes. Using 18000 evaluation episodes, we analyze performanceusing χ2 tests, logistic regression, survival analysis, and interaction modeling to assess successprobability, completion time, and data efficiency. Our results demonstrate that the sigmoidscheduler consistently outperforms linear and cosine schedulers, particularly as data availabil-ity increases, achieving higher success rates without lengthening task duration or inferencelatency. Given our setup, these findings highlight the impact of noise scheduling on diffusionpolicy performance, and emphasize its importance in model design choices.