Dynamic Initial Noise Generation for Diffusion-Based Robotic Manipulation Policy

Diffusion-based probabilistic models have achieved significant advancements in generative tasks, especially in image generation, and have recently been adapted for robotic manipulation tasks. However, their application to sequential action generation in robotics is hindered by the assumption of sample independence, which overlooks the temporal correlations inherent in manipulation tasks. Additionally, the conventional Pure Gaussian (PG) noise used in the reverse diffusion process often results in suboptimal performance due to discrepancies between training and inference phases arising from the variance schedule. To solve these issues, this paper proposes a novel Dynamic Initial Noise Generation (DING) method that leverages historical actions and environmental state differences to generate context-aware initial noise for the denoising process. Extensive experiments on the Push- T task demonstrate that the DING method improves success rates by approximately 5.03% compared to the PG approach, validating the effectiveness of our approach.