Interestingly enough, a car which simply refuses to drive in hurricanes, hailstorms, white-out snow, and so forth would be rather more than 20% superior to the average driver. So that's easily dealt with - if insufficient sensory data is available, either hand control over to the driver or, if possible, pull to the side of the road. The average driver likely does really horrible in large crashes and pileups, so merely the lowered reaction time might allow a huge improvement there (humans take a really long time to react, compared to computers). The average driver doesn't need to do evasive driving (as well as likely being horrible at it), so that's an unnecessary criteria. Similarly with tire eating potholes; even a tiny improvement would meet the 20% requirement you propose.
Ambiguous/contradictory road signs and unmarked roads are a navigation problem; given a sufficiently accurate database, as is likely to be present in most cities, they would be irrelevant (I expect that initial roll-out of driverless cars would occur in cities, where most traffic would be low-speed and highly structured. I may be wrong).
Unexpected car movements (due to snow, mud, or whatever else) I would agree to be fair things that need to be targeted, but I very much disagree with strong AI being needed for all of this. However, it's probably going to be quite a long while until the equipment is less expensive than hiring someone else to drive (which is a barrier for some group behaviors and sharing of data between cars). On the other hand, that's not necessarily a consideration which people factor in when buying extremely expensive cars.
One of the nice things about software drivers is that you can spend as much time as you want training a module for driving with, say, a given tire blown out and then swap that module in when actual conditions match the conditions that its trained for.
car which simply refuses to drive in hurricanes, hailstorms, white-out snow...
While admitted these conditions can sometime be anticipated, I think the point is that sometimes these conditions cannot be anticipated. You experience a sudden white-out a minute after a snow storm starts. What do you do? Breaking immediately is not the answer! (thus, the cars will need to drive...)
How Ironic. My very first job out of college, in 1995, was working for an avionics firm that was developing a system called "Tundra Tracker" - that would allow vehicles to navigate in whiteout conditions in the far north. I had the amazingly easy portion of taking the DGPS coordinates of maps/previous tracker runs and turning them into navigation maps. Challenges with the (dynamic) appearance of people, other vehicles, animals (Moose will ruin your day) crevasses, and other elements on the path were left to others to solve. :-)
Ambiguous/contradictory road signs and unmarked roads are a navigation problem; given a sufficiently accurate database, as is likely to be present in most cities, they would be irrelevant (I expect that initial roll-out of driverless cars would occur in cities, where most traffic would be low-speed and highly structured. I may be wrong).
Unexpected car movements (due to snow, mud, or whatever else) I would agree to be fair things that need to be targeted, but I very much disagree with strong AI being needed for all of this. However, it's probably going to be quite a long while until the equipment is less expensive than hiring someone else to drive (which is a barrier for some group behaviors and sharing of data between cars). On the other hand, that's not necessarily a consideration which people factor in when buying extremely expensive cars.