The third sentence of the fine article:

> The gap between simulation and real world remains large especially for perception problems.

They're training on video so their simulator would have to produce thousands of hours of realistic video. On top of simulating the aerodynamics and performance of the done itself. All to produce something which is going to be inferior to real-world data.

Does not need thousands, the article says they did only 40 hours of flying for this experiment at least

Human pilots also train using simulators first.

Well because if a human crashes a plane, that human doesn't stick around long enough to learn from that experience and do better next time.

Also the cost of a human crashing a plane is a bit more than a drone crashing itself, such that it's probably better to save on planes, and invest in simulators - whereas developing an accurate physics simulator for the purpose of training a drone might take more time/money than just letting it crash, and figure it out itself.

Type rating, yes, but initial training is usually done in a real airplane, at least here in the U.S.