this just feels like future state predictive reinforcement learning