One motivation for learning generative models of environments is to use them as simulators for model-based reinforcement learning. Yet, it is intuitively clear that when time horizons are long, rolling out single step transitions is inefficient and often prohibitive. In this paper, we propose a generative model that learns state representations containing ex