I'd bet it's a lot easier to write high reliability self driving software when you have multiple redundant sensors that don't all go to shit when it's foggy. There's probably a good reason that after a decade of r&d, every major player in this space with the exception of the one run by a narcissistic moron is using radar, lidar, and cameras in their sensor suite.
Humans can't drive safely in the fog. They can drive safely on a road travelled often even with low visibility by using visual hints they need to "scan" at lower speed. No car taking a cam approach for self-driving is doing that btw.
And yet, conceptually, there's still no capacity humans have that camera-only autonomous vehicles can't be given, even if they don't have it right now, right?
Theory says so, facts paint a different picture where CV systems are foiled more often (in good visibility conditions) than the same situation where a human could be foiled with camouflage or trompe-l'oeil.
Something is missing, time will tell if training will produce more solid models.
