Advantage being that if you can provide representative, simulated input then you can increase the training miles by orders of magnitude in the same amount of time, limited by computation rather than physical mileage.
Trying to use them to model real world scenarios would be useless in practice, due to the Ludic fallacy [0]. Real life is too complex to be modeled in any simulation.
As for modeling real life with simulations, the data exists for every type of accident a human has encountered. If you incorporate such into a simulation, plus randomly vary every free parameter, then your simulation will cover more scenarios than any human driver can possibly experience.
Thus, simulations should be able to help an autonomous vehicle outperform humans by a large margin, which is the only goal that matters.