RAAM: The Benefits of Robustness in Approximating Aggregated MDPs in Reinforcement Learning