Diffusion Model in Robotics: A Comprehensive Review

Abstract

Diffusion models are a powerful class of generative models that emerge in recent years to transform Gaussian noise into samples of the target distribution through an iterative denoising process. Due to the high training stability and powerful generative capabilities, diffusion models have surpassed previous generative models and demonstrate potential applications in the field of robotics. In the past few years, this area has gained increasing attention, and the number of studies applying diffusion models to the field of robotics has grown exponentially. This review aims to provide an overview of this emerging field, helping researchers understand the current state of development, with the hope of inspiring new research directions. First, we overview the foundation of diffusion models along with their development in recent years. On this basis, we provide an overview of the application of diffusion modeling in robotics from five aspects: scaling up robotics data, reinforcement learning(RL), imitation learning(IL), task planning and reasoning and other applications. We discuss and summarize the innovations, contributions, and limitations of these works. We then discuss the limitations and challenges faced by the field in terms of safety issues, real-time inference and model size, simulation to the real world gap, datasets and unified benchmarks and embodied foundation models. Finally, we summarize the review and provide an insight into future research directions.

Keywords: Diffusion Models, Robotics, Scaling up Robotics Data, reinforcement learning, Imitation learning, Task Planning