Another idea for improvements in the autonomous automobile space is the use of Generative Adversarial Networks (GAN’s) in the training thereof.

Normally in a GAN you might have one network attempting to identify birds in photographs, while the other network is doctoring photographs to get the first network to misidentify photos—either false-positives (birds where there are no birds) or false-negatives (non-identified birds).

In training autonomous vehicles, it may be possible to create virtual test environments where one network is working to successfully drive through simulations, while the second network is attempting to thwart that driving—either by causing the first network to crash its vehicle or by forcing the first network to halt its progress for trivial reasons.

I don’t know that anyone is doing this today, but it does seem a potentially fruitful avenue of testing and training.