That's a very interesting way of looking at it. Yes, you start with simulating something simpler than the real world. Then you use the real world. Then you need to go back to simulations for real-world things that are too rare in the real world to train with.

Seems like there ought to be a name for this, like so-and-so's law.