I just saw something recently where they said that most of the data taken by the autonomous cars on the road actually gets downloaded by hand and then shipped off to data centers in India for supervised learning. It's just too much data to transmit wirelessly or process on-board. Of course, if it was being processed on-board, that wouldn't be supervised learning.
Anyone that has tried to get a machine learning algorithm to learn to recognize things in an image either realizes how fragile the algorithms are or was never interested in noticing that in the first place. It's amazing how often they are right, but then again, they are not right that often. Scale invariance is a big problem, so of course people claim it a lot.