Learning Reward Functions for Robotic Manipulation by Observing Humans