Unsupervised Perceptual Rewards for Imitation Learning