Recent advances in diffusion-based embodied imitation learning: a survey

In recent years, embodied intelligence technology has held tremendous promise for unlocking the full potential of flexible, general, and dexterous robot systems, bridging the gap between virtual and physical intelligence. Imitation learning, a key framework in this field, enables the direct learning of policies from expert demonstrations, offering simplicity and efficient training. However, it faces challenges related to data efficiency, task modeling complexity, and generalization. Diffusion models, recognized for their solid theoretical foundation, strong distribution modeling capabilities, and stable training processes, have achieved notable success in image generation. Given their strengths in data generation, high-dimensional data processing, and complex distribution modeling, an increasing number of studies are integrating diffusion models into imitation learning to address these challenges.This article provides a comprehensive analysis of the applications of diffusion models in imitation learning, focusing on four aspects: principles, improvements, applications, and prospects. We first introduce the fundamental principles of diffusion models and explore their integration into imitation learning, forming diffusion policies, along with the basic framework of these policies. Next, we detail improvement methods for various components of diffusion policies, including conditional inputs, policy outputs, network architecture, and training strategies. We then review the applications of diffusion policies in robotic manipulation and mobile navigation. Finally, we discuss the challenges of using diffusion models in imitation learning and outline potential technical routes, offering references for researchers.