Future Tense

Avoiding the Uncanny Valley

New research suggests that robots that look almost like people are unsettling.

the humanoid robot shaped smartphone Robohon.
Sharp unveils the humanoid robot-shaped smartphone RoboHon at a preview of Asia’s largest electronics trade show CEATEC in Chiba, Japan, on Oct. 6, 2015.

Photo by Yoshikazu Tsuno/AFP/Getty Images

At 7½ inches tall, RoboHon is a walking, talking voice-controlled robot that can play games, dance, project visuals onto a wall—it’s even equipped with facial recognition software that allows it to get a feel for the mood you are in. The Japanese-made robot is set to hit shelves there in the first half of 2016, and it’s a clear sign that personal robots with social skills are on the horizon.

With a physical appearance that is loosely modeled after humans, RoboHon stands upright with two legs, two arms, wide-open eyes, and a welcoming facial expression. Its childlike voice and cheery demeanor instantly make you want to give it a big hug and be best friends forever. As more robots like RoboHon come into our everyday lives—as teachers, co-workers, aides for the sick and elderly, and even as companions—it is crucially important that they make humans feel comfortable interacting with them. If we fail to design robotic systems that can establish meaningful social bonds with people, then we will not be taking full advantage of the technology.

Hundreds of studies exploring ways to promote smooth human-robot interaction have been carried out in recent years. In general, scientists agree that as robots begin to appear more humanlike, people tend to respond to them more positively. These humanoid robots engage us better, since they can communicate in natural and familiar ways through social cues—like facial expressions, body postures, eye gaze shifts, and gestures. A machine with a lifeless appearance is much less inviting than one with a recognizable likeness, which implies familiar behavior.

However, the popular idea known as the uncanny valley suggests that there’s a problem with that approach. The hypothesis predicts that the positive relationship between a robot’s degree of human likeness and our affinity for it continues to grow only until a precise point. Specifically, when robots appear almost exactly human, people experience an unsettling feeling that causes revulsion. Something just feels “not quite right,” and the machine looks “creepy.” At this stage the positive relationship sharply turns negative, where it remains for a short period of time just before turning positive again when the robot starts to look completely indistinguishable from a human—a design feat yet to be fully achieved. (30 Rock has an excellent summary of the uncanny valley.)

The uncanny valley curve.

Courtesy of Maya B. Mathura and David B. Reichling

But the uncanny valley hypothesis, as put forth in 1970 by Japanese robotics professor Masahiro Mori, is just a logical prediction and not the result of objective experimental testing. British art curator and writer Jasia Reichardt later described it further in her 1978 book Robots: Fact, Fiction, and Prediction. Since its birth, the uncanny valley has lacked the detailed descriptions and rigorous explanations that are customary to most scientific hypotheses. Nevertheless, the paradigm has heavily influenced robotics design for decades.

Now that we’re getting closer and closer to designing robots that look like humans, testing the hypothesis is more important than ever. If the uncanny valley is in fact real, then trying our best to make robots that mirror humans exactly is a big design no-no.

Scientific investigation into the uncanny valley didn’t really start until about 10 years ago, which is roughly when researchers developed the ability to design highly realistic humanoid robots. One of the earliest studies to formally challenge the hypothesis, carried out in 2005, proposed that negative reactions to humanlike robots are more related to good or bad design aesthetics, and can occur at any level of realism. In other words, highly realistic and unrealistic humanoid bots can both cause revulsion with certain physical features—like bad skin, sickly eyes, significant asymmetry, and poor grooming. Conversely, clear skin, symmetry, and sharp grooming accompanied by oversize eyes, smaller-than-usual noses, or very large smiles can still be seen as highly attractive despite being unrealistic, if the proper balance is achieved. The study did find an uncanny valley effect when participants looked at a set of morphed images that followed a continuum from unrealistic to very realistic anthropomorphic robots, but that effect disappeared when the same images were made more attractive. The researchers concluded that although the uncanny valley may exist, careful design practices could help overcome it. A number of subsequent studies have both supported and conflicted with these results, and a recent systematic review described the empirical evidence for the hypothesis as ambiguous.

Despite the history of inconsistent research, a recent study in the journal Cognition provides compelling evidence to support the uncanny valley claim. What makes it different than earlier work is that researchers Maya Mathur and David Reichling may have avoided some pitfalls that could have obscured results in prior studies. In order to test whether the uncanny valley effect occurs with real-life robots, the researchers used 80 pictures of social robots that have actually been built, rather than using computer-generated morphed blends of human and robot faces, which often have unnatural distortions that cause strange features. They were also very picky about a lot of other factors—for instance, they only chose robots that were meant to interact with users (that is, not missing hair, facial parts, skin, or clothing), not marketed as toys (which are often made to look “adorable”), and capable of physical movement.

Wild-type robot face stimuli numbered and displayed in ascending
Real-world robot face stimuli arranged in order from very mechanical to very human.

Courtesy of Maya B. Mathura and David B. Reichling

In addition to rating how friendly each robot seemed, participants were asked to play an economic investment game with the bots to determine how much they trusted them. This is important because social trustworthiness is a big part of our willingness to interact with one another. The subjects were given up to $100 and were told to decide how much money to give to each robot in the hopes that they would receive a return on that investment.

The results showed a strong uncanny valley effect with both measures. Specifically, as robot faces appeared increasingly more human, their likability ratings increased up until they looked almost human, at which point the ratings dropped significantly, dipping down into the valley. Similarly, the amount of money wagered by participants first increased before drastically dropping, only to increase again when robots began to look identical to humans.

Uncanny Valley
Uncanny valley curves in a stimulus set (A) shown for likability (B) and trust-motivated wagering (C).

Courtesy of Maya B. Mathura and David B. Reichling

These findings have important consequences, not just for engineers and scientists, but also for anyone who wants to see robots come into our daily lives. Giving robots expressive physical features that allow them to communicate through social cues will make them more trustworthy, persuasive, and fun to be around.

At the same time, the results show that we must take careful caution not to fall into the uncanny valley by making humanoids that are too lifelike for comfort. Minor flaws and imperfections in appearance can give us a feeling of seeing dead matter impersonating humans, much like watching a zombie. This can elicit fear or disgust while reminding us of our own immortality. Fortunately, past research suggests that we can mitigate these effects by using good design practices that ensure that robots look as attractive and friendly as possible.

You might look at the new study’s data and conclude that we have already solved the uncanny valley design problem, since the most realistic robots reversed the valley with their high likability ratings. But the researchers used only static images of robots, so the participants couldn’t observe how those robots actually move. Since we are far, far away from being able to design machines that move realistically, even today’s most human-looking androids will likely still fall victim to the uncanny valley. Future studies that use dynamic videos of moving robots will be required to determine just how realistic robots would have to move to become likable once again.

We are social creatures, and as such require robots that are social as well. There is no doubt that getting their design completely right will require a lot of hard work and creative thinking. But with the many scientists out there currently picking apart these tough questions and applying that knowledge, we can be sure that we are well on our way to a society where interacting with cooperative, supportive, and incredibly productive robots is an everyday occurrence.

This article is part of Future Tense, a collaboration among Arizona State University, New America, and Slate. Future Tense explores the ways emerging technologies affect society, policy, and culture. To read more, visit the Future Tense blog and the Future Tense home page. You can also follow us on Twitter.