As AV's get very close to real world bulk testing I keep wondering what the safety goal for them should be - and who should set it. Statements like “anything to stop the daily slaughter on our roads" makes good headlines but no investment or safety engineering sense, (or even a moral sense to the family of somebody killed by faulty AV).
Going back in time when the Royal Aircraft Establishment developed auto landing in the 1960's they had a clear goal - to make auto landing ten times safer than a piloted landing in low visibility and they tested until they achieve their confidence limits on that target. I am sure there are learned articles and confidential company goals on AV safety but I can’t find any in the public media and you would like to think that “informed consent" would apply to AV safety too.
Setting a target is difficult and I would suggest only fatality goals are worthwhile, partially because injuries can vary hugely and also serious injury versus death is probably a function of the base car safety features rather than AV vs. human control.
So what target to set? You have to ask which country first. The USA has a road fatality rate of 1.16 per 100 million miles versus 0.2 in the UK so what might be an acceptable target in the USA would not reduce UK road deaths. The UK rate represents 816 fatalities in a population of 65 million. Pedestrian fatalities were 446 some of which could have been due to avoidable driver error. So a round number of 1,000 fatalities potentially avoidable with AV's might be sensible. Total vehicle mileage is 305 billion excluding trucks so one fatality per 300 million miles with human drivers.
So if human drivers are achieving 300 million miles per death what is a good AV target? Obviously they should be better or why spend all that money. The estimates are not going to be precise to some margin or multiple of 300 million is logical. The RAE used X10. Individual AV failures are less catastrophic than planes so maybe X 3 is OK. That gives nice round AV fatality target of one death per one billion miles traveled for the UK.
At that point the obvious problem arises - how can you verify a one in a billion failure and how many failures do you need to experience before you know you have reached your safety target of AV mass usage?
I am nor statistician or safety engineer so I have no idea how you do the testing but a) I think we need a clear target and b) that target is very, very low as human drivers are , in fatality terms in a country like the UK actually very safe when driving modern cars.