Almost-human is worse than clearly-not-human
As a robot's appearance is made progressively more human-like, comfort with it doesn't rise smoothly. It tends to increase for a while, then drop sharply as the robot gets close to, but not quite at, convincing human likeness, before potentially recovering if the resemblance becomes near-perfect. That dip is the uncanny valley: a robot that is obviously mechanical reads as a machine and is judged by machine standards, while a robot that is almost-but-not-quite human triggers the perceptual categories we use for people, and then violates them in small, hard-to-articulate ways, skin that doesn't quite move right, eyes that track a beat too smoothly or not smoothly enough, a face with slightly wrong proportions. The discomfort comes specifically from that category confusion, not from the robot looking artificial in an obvious way.
This matters for design because it means realism is not a variable you simply maximize. Pushing a robot's appearance further toward human likeness without also solving the fine-grained behavioral realism, facial micro-movement, timing of blinks and glances, natural pauses in speech, tends to make the mismatch more noticeable, not less, until the behavioral side catches all the way up. A team chasing visual realism for its own sake, because it seems impressive or advanced, can easily land a robot deeper in the valley than a simpler, more stylized design would have.