Playing Games in the Dark: An approach for cross-modality transfer in reinforcement learning