Agents need interactive worlds for reinforcement learning; high-quality simulations replace costly real-world data collection.
Train a web-navigation agent: a procedural simulator generates varied page layouts, bridge with human video demos for realistic clicking and scrolling, then evaluate transfer to real browsers where success rates exceed 85%.