PixL2R: Guiding RL Using Natural Language by Mapping Pixels to Rewards