Back testing however is possible. You could collect all recorded data from all driverless orgs, and require any new deploy be run over it. Plus a simulation.

For example, creating a standards body that collected this data and certified and release could be done by industry.

Unit testing is one technique used in autonomous vehicle verification, and that I think that includes the image classifiers and path planner, though I am less familiar with exactly how that is broken down, and how much is done in integration test like simulators vs unit test.

However, if you are saying that exhaustive unit testing of neural network based units is not possible, then I agree.