Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learning