Export Reviews, Discussions, Author Feedback and Meta-Reviews

This paper proposes a combination of deep learning and stochastic optimal control for control of non-linear dynamical systems from image inputs. The system learns a latent space representation of the true system state based on images of the state using deep variational auto encoders. Additionally, a locally-linear dynamics model is learned on the latent space which is used to generate state space trajectories by applying stochastic optimal control techniques directly on the latent space. The system is tested on four control tasks and the performance is compared with multiple baselines showing good performance. Quality: --------- The paper is logical and sound.