We tend to anthropomorphize objects and animals. That is, we project human characteristics like intelligence and emotion onto nonhuman things, especially if they exhibit some human traits. So, you'd think this would mean we would take to a humanlike android more readily than a metal mechanoid. We do apparently feel comfortable with robots that have increasingly human physical attributes up to a point, but once this point is passed, we are repulsed. This effect is called the uncanny valley.
The uncanny valley is a term coined by Masahiro Mori in 1970. To illustrate this idea, Mori created a graph with familiarity on the y-axis and human likeness on the x-axis and plotted our feeling of familiarity, or ability to identify, with various robotic forms or human representations. Industrial robots are near the origin, neither humanlike nor evoking a feeling of familiarity. Humanoid robots approach a peak, being both more familiar and humanlike. But after this peak, there is a sudden drop off into a valley (where things like corpses, zombies and prosthetic hands lie), and it rises again to a second peak as it approaches a living human. In his view, our comfort level increases as robots gain more human physical traits up to an as-yet-undefined point, at which the human traits suddenly render the robot unfamiliar and creepy. Both physical appearance and motion play a role, as unhumanlike motion can also throw something instantly into the valley.
Studies have borne out this idea -- and modified it a bit. Researchers Karl MacDorman, Robert Green, Chin-Chang Ho and Clinton Koch at the Indiana University School of Informatics used still images with the facial features and skin textures altered in various ways to garner participant responses. They found that eeriness levels were rated higher with faces that deviated from normal human proportions when the skin texture was realistic, but that backing off of the skin realism caused eeriness to decrease. These results seem to indicate that a mismatch between proportion and realistic detail might be a culprit.
A study by Ayse Pinar Saygin, Thierry Chaminade, Hiroshi Ishiguro, Jon Driver and Chris Frith used a moving robot (Repliee Q2, actually) to show that the uncanny valley effect might be caused by a disconnect between our expectations and reality in regard to the appearance and movement of an android. The researchers took functional magnetic resonance imaging (fMRIs) of participants as they viewed a series of videos of Repliee Q2, a mechanical-looking robot (the same android, but "skinned" to reveal its underlying parts), and a live human (in fact, the android's model), all performing the same actions. Participants' brains responded to the human and more mechanical looking robot very similarly. But when viewing the more humanlike android, different areas of the brain showed activity than in the other cases, and these areas had to do with connecting the visual cortex with parts of the brain that have to do with affecting and interpreting movement. It provides evidence that maybe the uncanny valley effect is triggered when something that looks mostly human moves in a non-human way (i.e., where appearance and movement do not match the way our brains think they should). Robots moving like we expect robots to move and humans moving like we expect humans to move, of course, do not creep us out.
One possible evolutionary reason for our revulsion to variances in the appearance and movement of an android is that any irregularity in a human form might indicate illness, and we may be hardwired to recoil to prevent the spread of disease. A certain otherness in another person might also trigger our aversion to people who we wouldn't deem as acceptable mating partners. But whatever the underlying biological reason, roboticists are looking for ways to keep their creations out of the valley.