Well I was a little confused on that point. I think that it is a kind of stitching but maybe not at the overlapping pixel level the level the way many stitching algorithms work. Note that Karpathy made a point that it's essential for them to calibrate out the slight alignment variances in each car (he made the amusing comment "all of our cars are slightly cockeyed, in a slightly different way") and he did say the images are "stitched up").The recent AI Day presentation demonstrated they do have cross camera object recognition (recognizing objects that straddle two cameras) instead of relying on stitching. on a prr-car
I would certainly agree that more overlap, and less distorted distorted overlap, would help a lot in in creating the continuous panoramic view more quickly and confidently. And I've been a big proponent of more and better camera angles. However it's also true that our brains are very talented at filling in a world view from fairly distorted and low-resolution peripheral vision, based on anchor objects rather than a meshing 9f fine-grained detail. So in the NN, It may not be a case of being presented with a nicely stitched 360 video feed, hut rather a poorly-stitched and and somewhat imperfect panorama that still contains the elements necessary to extract the needed features.