The combination of deep learning processing with traditional underlying dynamical systems for planning has been challenging, but it can be achieved by choosing a latent variable that informs about the future before taking high level action, which leads to faster learning with better credit assignment.