Enhancing Vision-Language Model Training with Reinforcement Learning in Synthetic Worlds for Real-World Success