>they would train the pipeline on a raw-ish image data to have better consistency? I would think that some kind of algorithm "bettering" the image

of course, whatever "bettering" algorithm makes sense, the neural network would probably emerge it inside itself as a result of training. My comment was more about hardware in the sense that "garbage in - garbage out", and so you'd want to have a decent hardware which is available and pretty cheap today while Waymo seems to be using, i'm stating it politely :), very-very-commodity level.

>Also, what looks good to us may not be "good" to a machine.

the kernels of the first layers of visual CNNs (at least the ones i looked at) converge to the Gabor-like kernels what are in the first layers of the human visual cortex. Bunch of years ago, before deep-learning CNN, etc. i saw an article (in neurophysiology field) proving an optimality of such kernels for those visual tasks. Of course things may and probably diverge a lot in the deeper layers, yet deep dreaming, etc. suggests to me that the nature of the divergence may be less severe than expected.