You can get useful levels of robot training out of Super Mario Bros 2; see John Carmack's talk from last year: https://www.youtube.com/watch?v=4epAfU1FCuQ

When your training simulation fails to reproduce certain features of the real-world environment, your control model is likely to have trouble with them, so more realistic simulations are helpful, but even fairly unrealistic training simulations are pretty useful.

In general, training neural networks to control robots is going to require simulations, ideally differentiable simulations. Even if they aren't differentiable, though, you can SGD your simulated training runs across a data center in a way that you just can't with training runs on Roombas.