The problem with simulator testing is that you can't test scenarios that you didn't think of. This is particularly important to find problems arising from multiple simultaneous situations. For example, you might test the scenarios "front camera obscured by rain", "car ahead of you performs emergency stop", and "dog runs into street", but that doesn't necessarily tell you how the car will respond to a combination of the three.
Real life is far more creative than any scenario designer.
Which is why you should do both. A simulation can test millions of permutations -- including arbitrary combinations of events, and in far more variety than could be tested in a reasonable amount of time on real roads -- and can verify that software changes don't introduce regressions. Real-world testing introduces an element of randomness which provides additional insights for the simulation test cases.
Ultimately, governments should probably develop their own simulators which run the autonomous car through a large battery of scenarios, including scenarios which include disabling some of the car's sensors. Then autonomous vehicles from different manufacturers could be validated on a standard test suite before being allowed on the roads, and when real-world incidents occur in which an automated car makes a bad decision, those incidents can and should be replicated in the simulator and all certified vehicles tested. They should also do real-world testing, but I suspect that in the long run simulations will provide much greater confidence.