Denoising diffusion probabilistic models (DDPMs) are a class of models that can learn to conditionally generate data from random noise. This is achieved by iteratively applying a noise function to the training data and subsequently training the DDPM to learn the inverse process that can iteratively denoise the data.
For a few years, DDPMs have been successfully deployed for generating high-quality images [1, 2]. More recently, research has focused on applying these models to robotic applications [3, 4, 5]. This thesis will expand on these efforts to exploit the potential of DDPMs for robotic planning and control.
The main aim of this thesis is to investigate how to introduce physics-informed priors into the DDPMs. Adding information about the physical domain should increase its sample efficiency by guiding the model to physically plausible solutions. Additionally, we would like to implement the control scheme with the DDPM on a real robot.
