that's the thing, at the point that the computer is already that good, missing that one accident isn't optimal sure, but it's definitely better than the current state of vehicle operation.
An autonomous driver which was twice as dangerous as a human, then, would still go 37k to 50k miles between collisions. A human who was trying to backstop that robot's fallibilities would be required to pay close attention for weeks between actions. Which is inhuman.
Source: I worked for an insurance company for 8 years, and part of my job involved updating casualty models.
There are a LOT of accidents, but most are fender-benders. I guess if you exclude those, yes, a serious collision occurs about those numbers.
In the US:
The Department of Transportation gets reports of one accident per 250k miles (roughly). It is broadly agreed that many accidents are unreported, with estimates of the true rate ranging from 1/200k miles to about 1/75k miles.
For example, this document from the US Department of Transportation:
https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/...
suggests 5,687,000 total crashes (including fatality, injury, and non-injury) in the US in 2013.
This graph: https://fred.stlouisfed.org/series/M12MTVUSM227NFWA
puts total vehicle miles driven in 2013 in the US at 2,980,181,000,000.
So divide: 5,687,000 / 2,980,181,000,000
That's something we've discussed a lot here - there's NO way self-driving cars can go around South American streets - unless they learn to be very aggressive, beep the horn, cross streets whenever they can, shout and otherwise interact with other drivers.
And we mostly don't have highways. Americans drive a lot in highways, that must skew the per mile accidents.
I don't know how often accidents like fender-benders go unreported in the U.S. though.
What I wanted to mention is that there are a LOT more cars it the U.S., and the average US driver drives a LOT more than the average Uruguayan driver. (I'd have to look up hard numbers, but that's the gist of it)