Imaging a dog trained to drive a car. It would probably be very good at visual tasks and presumably even excel at some auditory task with its superior hearing. However, some things will be a lot harder for the dog, such as recognizing traffic signs and applying it to its driving.

Many things will be completely  impossible  for the animal, i.e. reading and fully understanding language from signs or in spoken communication. Many other, maybe most,  safety critical task will also be impossible; estimating when the car will run out of fuel if it is filled up with 50 liters or ”I see a school so there might be children behaving irrationally” would be other examples. In general any task requiring reasoning of the form A may-lead-to B and B may-lead-to C, will fail, while this kind of reasoning turns out to be the way humans solve many safety critical task in autonomous driving. If there is one ability that separates humans from animals, it would be that humans has the ability to reason to think and communicate symbolically, to ponder i.e. infinite amounts or plan long ahead, or to deduce facts around never before seen scenarios.

Reasoning is paramount to safety. Ideally you would like to be able to reason to give mathematically true and meaningful boundaries to future uncertainties.  Quantifying risk with a mathematical proof is the safest way to go.

One challenge for the researcher and developer is that there are not many methods in the research community suitable for uncertainty reasoning with the right rigor and at the same time the required “creativity”.

However, we have good experiences with a method called genetic reasoning where reasoning paths are evolved in simulated evolution to obtain proofs on safety in quantitative domains.

This kind of reasoning algorithm would be the top most layer in any successful autonomous driving application - we for sure need; reactive behaviour as well as model building approaches but without a top level “dry” reasoning layer, it will be at best as an animal driving a car, which is not good enough for safety.