Towards verifying neural autonomous systems