Learning Representations via a Robust Behavioral Metric for Deep Reinforcement Learning