Crossing the uncanny valley

ROBOT-MAKERS and the animators who design characters for films and video games face a paradox. People readily accept machines and cartoons that are simplifications or distortions of the human form. Simulacra that are intended to look like real people, though, are frequently perceived as creepy. In November 2004, for example, two films intended as light entertainment were released to very different receptions. “The Incredibles”, a cartoon in traditional style, was one of the most successful movies in history. “The Polar Express”, which used motion capture and computer graphics to produce an animation whose characters looked almost human, received a critical panning: one reviewer suggested its characters were so frightening the film should be subtitled “The Night of the Living Dead”.

The difference lay in that one crucial word: almost. Workers in the field refer to the perceptual crevasse which separates acceptable caricature from accurate representation as “the uncanny valley”—and the “The Polar Express” fell right into it. Mapping the uncanny valley, to avoid its perils, would be of great benefit to film-makers. It would also, as robots become smart and safe enough for use outside factories, help engineers to design plastic pals who are truly fun to be with.

Though several previous expeditions have been lost in the valley, there is no shortage of volunteers to have another go. The latest pair to harness up the metaphorical huskies are Chin-Chang Ho and Karl MacDorman of the Indiana University School of Informatics. They think they can find their way by following a new compass direction—the quality of eeriness.

The valley below

The idea of the uncanny valley was originally proposed by Masahiro Mori, a Japanese roboticist, in 1970. Though he had no hard data, his intuition was that increasing humanness in a robot was positive only up to a certain point. Dr Mori drew a graph (see chart) with “human-likeness” on the horizontal axis and a quality he called shinwakan (variously translated as “familiarity” and “comfort level”) on the vertical one. As an object or image looks and behaves more like a human, the viewer's level of shinwakan increases. Beyond a certain point, however, the not-quite-human object strikes people as creepy, and shinwakan drops. This is the uncanny valley. Only when the object becomes almost indistinguishable from a human does shinwakan increase again.

Dr Ho and Dr MacDorman accept the general idea, but they began by throwing out the idea of shinwakan. In their study, just published in Computers in Human Behavior, they say that Dr Mori's ideas of familiarity and comfort level do not properly get at the quality of uncanniness. Neither do some suggested alternatives, such as warmth and likeability. The wicked queen in Disney's “Snow White”, for instance, was hardly likeable. But she was not uncanny either.

To isolate the factors that really affect how people feel about simulacra, the two researchers rounded up several hundred undergraduates and showed them ten video clips, five of robots and five of animations. These included sequences from “The Polar Express”, “The Incredibles” and also an animation of Orville Redenbacher, an American businessman who died in 1995, that many people think falls right at the bottom of the valley. The robots were the Roomba, a disc-shaped autonomous vacuum cleaner, and four anthropoid machines of varying degrees of humanness.

The volunteers were asked to apply ratings from dozens of scales to each video: machinelike to humanlike, synthetic to real and so on. Scales that turned out to measure the same qualities with different words were eliminated and the researchers eventually lighted on 19 that described aspects of four underlying qualities that they dub attractiveness, eeriness, humanness and warmth. According to Dr MacDorman, all four are important qualities for designers. A robot that exhibits warmth and attractiveness will be easier to interact with than one that looks cold and ugly. Only two of them, however, are needed to explain the uncanny valley. These are humanness and eeriness.

Eeriness is not quite the same thing as comfort level, likeability or even strangeness. Levels of eeriness were indicated by eight descriptive scales, including “ordinary/supernatural”, “boring/shocking” and “uninspiring/spine-tingling”. By plotting perceived humanness along the horizontal axis and eeriness along the vertical, Dr MacDorman says that he can recreate Dr Mori's chart of the uncanny valley, this time using real data about how people feel about a particular robot or animation.

That could be useful information. Although robot butlers remain a distant dream, if people are ever to interact with robots in a comfortable way those robots will need to avoid making users' skins crawl. In the meantime, the video-game industry is continually trying to increase the realism of its basketball players and brawny mercenaries. Hollywood, too, would probably enjoy replacing stuntmen—and perhaps even temperamental stars—with computer-generated versions. A world without celebs? That really would be eerie.