I. Introduction
Variational latent variable methods are widely used in the machine learning community for automatically capturing major factors of variation into a low-dimensional latent space. In robotics, low-dimensional factors of variation have already been identified for many common dynamics phenomena: inertial characteristics, friction coefficients, spring elasticity constants, etc. These are the key adjustable parameters in many general-purpose simulators, and can be viewed as a highly effective ‘latent space’ representation. Our goal is to ensure that scalable variational methods can leverage this representation and can incorporate hardware observations in a data-efficient way to close the sim-to-real gap.