Zero-Shot Transfer with Deictic Object-Oriented Representation in Reinforcement Learning