Keywords

Abstract

Probabilistic modeling provides the foundation for learning under uncertainty, evolving from classical methods like Gaussian mixtures and Bayesian networks to modern deep generative frameworks. Diffusion models represent a breakthrough by learning data distributions through a gradual denoising process, offering stability and expressiveness. Extending this idea, Diffusion Policies apply diffusion-based generative modeling to robotic control—enabling robots to learn rich, multimodal action distributions instead of deterministic commands, leading to more robust and adaptable behavior.