Mutual Information Maximization for Robust Plannable Representations