This is true, to the extent that the selection of input parameters to simulate provides coverage over the domain of all possible input parameters. At some point you're back to testing the failure of imagination of the simulation/test creators.

The problem is the simulation by definition can only simulate things that are accounted for. Any number of completely arbitrary out of the blue things can happen in real life.

Simulations can make use of random number generators and could in theory, following the lead of a project like afl, adaptively find algorithmic weak spots.

Simulations are only useful up to a point.

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.

[0] https://en.wikipedia.org/wiki/Ludic_fallacy

I don't think the connection to the referenced fallacy is nearly strong enough to serve as a QED on its own. It's also seemingly promoted only by one person.

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.

Oooh, adversarial scenario generation! I think there's something to that.