I find it pretty clever to generate Robot faces rather than whole robots or other pictures. Human brains are optimized for face recognition, which is why we can tell even minor differences in faces far easier than minor differences in other pictures.

Therefore, generating human faces would serve this purpose even better, but those are at risk of falling into the uncanny valley. (http://en.wikipedia.org/wiki/Uncanny_valley)

Using clearly non-human faces (such as Robot faces) avoids that problem.