DARLA: Improving Zero-Shot Transfer in Reinforcement Learning