[4O3-OS-16e] OS-16

[4O3-OS-16e-01] Consistency Models based Scalable Diffusion Policy

〇Sodtavilan Odonchimed¹, Yuya Ikeda¹, Ryosuke Takanami¹, Tatsuya Matsushima¹, Yuta Oshima¹, Takuya Okubo¹, Kai Nabeta², Yutaka Matsuo¹, Yusuke Iwasawa¹ (1. The University of Tokyo, 2. University of Tsukuba)

Keywords:Diffusion Models, Imitation Learning

We introduce Consistency Policy, an imitation learning model that can operate in real-time. In imitation learning, the input is required to be multimodal and the output is multimodal in nature, making it more difficult than general supervised learning. In this case, the acquisition of action distributions using generative models, especially Diffusion Policy, has reached a high level of accuracy. However, Diffusion Policy has a trade-off between accuracy and real- time performance. In this study, we propose a new policy model for robot learning that can operate in real-time while maintaining the accuracy of Diffusion Policy: Consistency Models, which have sample speed performance that Diffusion Models cannot achieve, and Conditinoal Consistency Models, which can perform multimodal conditioning. Consistency Policy has achieved in a 10-times increase in action generation speed compared to the Diffusion Policy.