I think solving an "easy" problem is a good step on the way to a hard one, as long as the tool is set up to allow you to get out of a local maximum. Clearly there are locations where self driving is a smaller problem space than others, and I don't think it's bad to solve for those locations first.
I don't think it would be a fail if self driving cars were able to only operate in sunny locations with good roads. The caveat is that the outcome shouldn't just be a "magic" ML model that can't be modified to handle rain or potholes, it should be a set of tooling that allows you to make ML models that solve a variety of problems.
So build a covered test road for your initial trials - then unleash them on real road situations. I'd be much more ready to trust a self-driving car that was born and raised on Boston streets than one in the wide, luxurious and pristine SoCal streets.
I'm not saying you need to teach your kid to swim by throwing them into the sea with weights tied around their ankles - you can ramp up to that... but you need to start exposing things to real world conditions pretty early on in development lest something, like a LIDAR censor close to the road surface in the front of the car, force you into a huge redesign when you discover that sheets of slush and road salt will liberally coat every front-facing surface of your vehicle driving in Boston in the winter.
I don't think it would be a fail if self driving cars were able to only operate in sunny locations with good roads. The caveat is that the outcome shouldn't just be a "magic" ML model that can't be modified to handle rain or potholes, it should be a set of tooling that allows you to make ML models that solve a variety of problems.