That clearly is the case, it was my experience growing up that even seasoned New England drivers couldn't take the weather seriously and ended up in a ditch on the side of the road.
That's not really a good point. When New England/Canada drivers drive through a snowstorm, the vast majority of them get to where they were going just fine.
In comparison, basically no self driving car could safely drive 100 ft during a snowstorm.
Getting self driving cars to be as good as humans during harsh weather (which is not that good I'll give you that) will take 10 years in my opinion (I say this as a grad student in mobile robotics).
I'd say betting any tech 10 years down the line is a fools game. in 2008, people were saying the same thing about object detection, or literally 100s of other things that deep learning has done that people said was impossible.
Autonomous cars are already safer on ice. being able to individually apply abs to each wheel in micro second speeds of detection of change in terrain is far more accurate than any human.
As far as snowfall, for the vast majority of places, hi def maps and semantic knowledge of signage already exists, so as long as the car can localize, it can obey snow covered signs, and google and others are already working on filtering out weather "noise".
I think you're mostly right; the only thing I'm more skeptical about is the availability and reliability of high def maps and semantic knowledge as you say. Maintaining such thing will be expensive and I wonder how it will scale country-wide as opposed to a few select cities. I bet using the self driving cars themselves to do the mapping and correct mistakes will be a big part of that.
Rough localization in a snowstorm (as in good enough to use prior knowledge of signs) or rain shouldn't be too hard. GPS still works in bad weather and lidar localization (scan-to-map matching) should be hampered but still usuable.
