This week I looked at a paper by Bartneck, Kanda, Ishiguro and Hagita discussing whether the concept of the uncanny valley which originated in the 1970s is actually true. Since we now have extremely human-like robots available the question is how do we actually perceive them and is it possible to come out of the uncanny valley?
A version of this paper is available online at http://www.bartneck.de/publications/2007/uncannyCliff/.
Bartneck, C., Kanda, T., Ishiguro, H., & Hagita, N. (2007).
Is The Uncanny Valley An Uncanny Cliff? In RO-MAN 2007 – The 16th IEEE International Symposium on Robot and Human Interactive Communication (pp. 368–373)
The uncanny valley theory proposed by Mori in 1970 has been a hot topic in human robot interaction research, in particular since the development of increasingly human-like androids and computer graphics. In this paper we describe an empirical study that attempts to plot Mori’s hypothesized curve. In addition, the influence of framing on the users’ perception of the stimuli was investigated. Framing had no significant influence on the measurements. The pictures of robots and humans were rated independently of whether the participants knew a particular picture showed a robot or human. Anthropomorphism had a significant influence on the measurements, but not even pictures of real humans were rated as likeable as the pictures of humanoids or toy robots. As a result we suggest the existence of an uncanny cliff model as an alternative to the uncanny valley model. However, this study focused on the perception of pictures of robots and the results, including the suggested model, may be different for the perception of movies of moving robots or the perception of standing right in front of a moving robot.
This paper looks at the premise that the original work done in the 1970s that described the concept of the uncanny valley is not accurate and that now there are better ways to measure it. It proposes that actually the original work has often been translated with familiarity versus human likeness but that this is a mis-translation of the original Japanese and that it should actually be affinity or likeness (this was later confirmed when the paper was formally translated in 2012) and as such they will be looking at the concept of likeability rather than familiarity in this paper. In order to measure this they also look at the psychological concept of framing – where pre-existing information is available humans will attempt to determine what will happen next and what to do if the expectation is unfulfilled (an example given is walking into a restaurant). In this context the framing is whether the image supplied is framed as a human or a robot. If a robot is framed as a robot but seems more human then it may end up with a greater likeability score than a robot framed as a human in the first place.
Considering the difficulties that can exist assessing this area rigorously and scientifically the team attempted to construct a series of experiments that would remove as many human biases as possible. The outcome was ultimately an 36-question test that a participant would have to answer with the questions appearing in a random order. For each of 18 pictures there was one question asking for the likeability and another asking how human-like the picture was. The human-like scoring was asked in such a way that each participant had to answer a number of questions from a 7-point in an attempt to standardise the responses. The picture types that were assessed were real humans, manipulated humans (skin tones made more android-like), computer generated images, androids, humanoid robots and pet robots. Each picture was categorised within a framing category – as either robot, human or simply face. With the exception of humanoid robots and pet robots (which were clearly robots) all pictures were assigned to a framing category at random (so that an android could be framed as a human or a human as a robot for example). This framing exercise occurred three times resulting in three different sets of questions each containing 36 questions on the same 18 pictures but framed in different ways. All pictures were of face-only, female-looking images to remove gender biases and context of an image. The female gender was chosen as no male android robots were available at the start of this study. The Cronbach’s alpha score for the results indicated a “Good” internal consistency suggesting that the controls put in place for the experiment worked as intended.
From the results there were two key outcomes:
- Framing had no significant impact on the results
- The level of anthropomorphism had significant influence on human likeness and likeability
Interestingly when an image was framed as either human or robot it made no difference to how people overall perceived the image. The level of anthropomorphism in an image (how realistically human it was) did, however, have a significant impact on both scores. In many ways values backed up the original work by Mori – toy robots that didn’t try to emulate being human actually got greater likeability scores than real humans but interesting rather than an uncanny valley the peak expected on the right of the valley dip is much lower – so much so that humans actually come out as not being liked as much as toys and computer generated imagery came out more human-like in some cases. The results were broken down to the level of per image to show differences when it came to individual images but the results do support that this shallow quadratic curve rather than the traditional view of a valley that rapidly becomes more likeable and human.
A lot of work was undertaken to try and give quantitative results to an area that is difficult to study. The authors did pick up on a couple of areas for critique – namely Japan all participants were Japanese which is a knowingly pro-robot culture and that ti is an intrinsically difficult question to test for. The authors go on to suggest further work around framing and adding in a larger variety of samples specifically looking at moving videos rather than static images.
I found this paper really interesting as it throws the question of whether there is any validity in trying to create ever-more human robots so that people feel comfortable interacting with them. All the evidence collected in this paper suggests that less-human looking robots could actually result in being more liked by human users which is the ultimate reason to develop robots.