If you invest (significantly) into a test suite that realistically mirrors the real world, that argument doesn't fly. Think Chaos Monkey squared. But it's neither cheap nor fast. Sure, some rare edge cases are probably impossible to model even with great care. But that's not the stage self driving is at, they struggle with pretty common and predictable problems.
You’re talking out of both sides of your mouth here. Even the greatest test suite on the planet will miss edge cases in the real world. Self driving cars are already 99.999% successful and effectively perfect in simulation.