Keywords: Representation learning for planning, Unsupervised learning, Approximate inference via stochastic optimal control
TL;DR: We propose a novel stochastic optimal control-based framework for representation learning and planning of dynamical systems.
Abstract: We present a novel framework for representation learning that builds a low-dimensional latent dynamical model from high-dimensional sequential raw data, e.g., video. The framework builds upon recent advances in amortized inference methods that use a differentiable network to output samples from a variational distribution given observations as inputs, and takes advantage of the duality between control and inference to approximately solve the intractable inference problem using the path integral control approach. We also present an efficient planning method that exploits the learned low-dimensional latent dynamics.