Trajectory generation and control of dynamical systems via Conditional Denoising Diffusion Probabilistic Models
Please login to view abstract download link
Denoising diffusion probabilistic models (DDPMs) are a class of (deep) probabilistic generative models learning generative data distributions out of which we can sample new data. DDPMs generate data from random noise by performing a stepwise denoising of the random vectors. The denoising process, often referred to as the reverse process, is learned from data with the goal of reverting the forward process of gradually adding noise to the data. In this talk, we show (i) how DDPMs can be used in the context of dynamical systems for generating accurate and data-consistent trajectories over variable-length prediction horizons, and (ii) how the generation process can be conditioned to different initial conditions, parameters, or control inputs.
