Rolling Diffusion Policy for Robotic Action Prediction: Enhancing Efficiency and Temporal Awareness

Chanhyuk Jung, Dasom Ahn, Sangwon Kim, In-su Jang, Kwang-Ju Kim, Sungkeun Yoo, Byoung Chul Ko

Keywords: Diffusion policy, Visuomotor policy

TL;DR: We propose Rolling Diffusion Policy, a novel approach that accelerates diffusion-based robotic control by reducing iterative steps through a dynamic rolling mechanism, achieving faster inference while maintaining high-quality action generation.

Abstract: Diffusion models have shown strong potential for generating high-quality robotic action sequences, yet their iterative nature often incurs substantial computational cost. In this paper, we propose a novel Rolling Diffusion Policy (RDP) that accelerates diffusion-based control by reducing the number of iterative steps required for action generation. Our approach introduces a dynamic rolling mechanism that incrementally refines action trajectories while effectively capturing the temporal dependencies inherent in robotic systems. Integrating this mechanism into the diffusion policy framework enables faster inference while maintaining high performance. Extensive experiments on simulation benchmarks reveal that RDP achieves comparable or improved performance compared to conventional diffusion-based methods, paving the way for real-time applications in complex robotic environments.

Submission Number: 1