In January 2026, a team at Columbia Engineering published a paper in Science Robotics announcing that they’d built a robot capable of learning realistic lip movements by watching its own reflection and studying human videos. The robot could speak in multiple languages and sing. Hod Lipson, the lab’s director, framed the significance plainly: “There is no future where all these humanoid robots don’t have a face. And when they finally have a face, they will need to move their eyes and lips properly, or they will forever remain uncanny.” His co-researcher Yuhang Hu added: “We are close to crossing the uncanny valley.”
That phrase—uncanny valley—has been floating around robotics and cognitive science for over fifty years, and it describes a problem that gets more commercially urgent every quarter. Tesla, Figure AI, Agility Robotics, Boston Dynamics, and a growing roster of companies are building humanoid robots designed to operate in spaces built for humans—warehouses, hospitals, factories, homes. The machines are getting better at walking, grasping, navigating, and following instructions. The question of whether people will actually want to be around them is a different engineering challenge entirely, and it’s one that can’t be solved with better actuators.
The graph that launched a thousand nightmares
Japanese roboticist Masahiro Mori proposed the concept in 1970 in an essay for the journal Energy. The idea is deceptively simple: as a robot becomes more human-like in appearance, people’s emotional response becomes more positive—up to a point. A cartoon robot is charming. A robot with a humanoid shape and some facial features is engaging. But as the resemblance approaches near-human levels without quite getting there, the response doesn’t just plateau. It collapses. The emotional curve drops into a trough of revulsion, unease, and what Mori called bukimi—a Japanese word that translates roughly to “eeriness.” That trough is the uncanny valley.
Mori’s original graph wasn’t based on experiments. It was based on his personal observations and intuitions, which is a detail that tends to get lost in the retelling. The essay was more philosophical provocation than empirical finding, and it sat in relative obscurity for decades before roboticists and CGI animators in the 2000s rediscovered it and realized they’d been stumbling into the valley independently. The 2004 Robert Zemeckis film The Polar Express—in which Tom Hanks was motion-captured into a CGI character that audiences found deeply unsettling despite technically impressive animation—became the canonical example of uncanny valley in popular culture. The technology was extraordinary. The result was a children’s movie that gave children nightmares.
What makes the uncanny valley genuinely interesting, rather than just a fun piece of trivia to reference when a new humanoid robot demo goes viral, is that nobody fully agrees on why it happens. And the competing explanations have very different implications for whether it can be solved.
The competing theories
The most commonly cited explanation is perceptual mismatch—the idea that human brains are optimized over millions of years of evolution to process human faces with extraordinary precision, and when something is close to human but slightly off, the discrepancy triggers an error signal. We’re wired to detect subtle abnormalities in faces because historically, detecting disease, deception, or unfamiliarity in the people around you was a survival advantage. A robot that’s 95 percent human-looking trips the same alarm system that would fire if you encountered a person with something wrong with them—asymmetrical facial movement, delayed eye tracking, a smile that doesn’t reach the upper face. The problem isn’t that the robot looks bad. The problem is that it looks almost right, and the remaining five percent registers as pathological.
A second explanation focuses on categorical ambiguity. Things that sit cleanly in one category—”obviously a machine” or “obviously a human”—are psychologically comfortable because the brain knows what schema to apply. Things that fall between categories—not quite machine, not quite human—create cognitive dissonance. You don’t know whether to treat it as an object or a person, and that uncertainty is inherently aversive. This explanation has roots in psychological research on disgust responses to category violations more broadly—the same mechanism that makes chimeric creatures in horror films effective.
A third theory, supported by a 2021 study published in Computers in Human Behavior, argues that the uncanny valley is driven specifically by the perception that a robot has feelings. Researchers found that humanoid robots are unsettling in part because people unconsciously attribute the capacity for subjective experience—what philosophers call phenomenal consciousness—to things that look human. When you look at a robot face and your brain automatically assumes something is going on behind those eyes, and then another part of your brain recognizes that nothing actually is, the collision between those two assessments produces the eerie feeling. The study demonstrated that “dehumanizing” a humanoid robot—explicitly telling people it has no feelings—significantly reduced uncanny valley responses, and this effect held up in a field study with hotel guests interacting with real robots in Japan.
This third theory is the one with the most interesting commercial implications, because it suggests the uncanny valley isn’t purely a design problem. It’s a framing problem.
What the latest research shows
A February 2025 study from researchers working with Nadine, a hyper-realistic humanoid robot, tested whether equipping a robot with LLM-powered conversational abilities could reduce uncanny valley effects. Eighty participants interacted with the robot and completed pre- and post-interaction surveys. The findings were promising and limited in roughly equal measure: LLM-enhanced conversations significantly reduced feelings of eeriness and increased perceptions of pleasantness and approachability. But—and this is the part that matters for anyone trying to commercialize these things—a subset of users continued to report discomfort even after extended, high-quality conversation. The uncanny valley effect shrank. It didn’t disappear.
The regression analysis revealed something counterintuitive: conversational naturalness and interestingness predicted willingness to continue interacting, but visual human-likeness did not. In other words, once the conversation was good enough, how human the robot looked stopped being a significant factor in whether people wanted to keep talking to it. The implications of that finding, if it replicates, are substantial—it suggests that investing in better AI conversation may be a more efficient path past the uncanny valley than investing in more realistic skin texture.
The Columbia Engineering lip-movement paper approaches the same problem from the opposite direction. Their argument is that facial expression—particularly lip synchronization during speech—is so fundamental to human communication that getting it right is non-negotiable. Nearly half of human attention during face-to-face conversation is directed at the speaker’s lips. A robot whose mouth movements don’t match its speech triggers uncanny valley responses even if everything else looks perfect, because the mismatch between what you hear and what you see is processed as a deep wrongness. Their robot learned lip movements through self-supervised learning—watching itself in a mirror and comparing its movements to human video—rather than being explicitly programmed. The result was lip synchronization natural enough that Lipson, a self-described “jaded roboticist,” reported involuntarily smiling back at his own creation.
The design choice nobody talks about
Here’s the thing about the uncanny valley that tends to get overlooked in favor of the more dramatic question of “can we cross it?”: most of the companies actually building commercial humanoid robots have decided not to try.
Look at the design language of the machines that are closest to deployment. Boston Dynamics’ Atlas has no face at all—it’s a headless torso with extraordinary agility. Agility Robotics’ Digit has a face that is deliberately stylized and cartoonish—two large “eyes” on a rounded head that reads unmistakably as friendly and unmistakably as not human. Figure’s robots have a visor-like head that evokes a motorcycle helmet more than a human face. Tesla’s Optimus has gone through several iterations, and each one has moved further from human facial features toward a smooth, featureless faceplate.
These aren’t failures of engineering ambition. They’re deliberate design decisions informed by the uncanny valley research. If your robot’s face can’t be indistinguishable from a human face—and no current robot face can—then the safest design choice is to make it clearly, obviously, comfortably non-human. Stay on the left side of the valley. Don’t attempt the crossing. The cartoon robot, the friendly geometric face, the abstract visor—these designs maintain positive emotional responses without risking the plunge into eeriness.
The companies pursuing hyper-realistic human faces—Hanson Robotics (creators of Sophia), Hiroshi Ishiguro’s Geminoid series—tend to be research labs and publicity vehicles rather than commercial deployment operations. Sophia is famous, but Sophia isn’t picking items in a warehouse or delivering medication in a hospital. The robots that are actually being deployed at scale are the ones that look like robots, and that’s not a coincidence.
Whether any of this matters
The pragmatic argument—increasingly popular among robotics engineers who are tired of the uncanny valley being treated as an unsolved existential crisis—is that the entire question is overrated for most commercial applications. A warehouse robot doesn’t need a face. A manufacturing robot doesn’t need to be likable. Even in healthcare and hospitality, where robots interact directly with humans, the evidence suggests that functional competence matters more than facial realism. The Japanese hotel study found that guests’ satisfaction with robotic service was unaffected by whether the robot was framed as human-like or machine-like, as long as it performed its job.
The counterargument—and it’s a serious one—is that the applications where the uncanny valley matters most are precisely the applications where humanoid robots would generate the most value. Eldercare. Companionship. Therapy. Education. Customer-facing roles that require trust, rapport, and emotional connection. If you want a robot that an elderly person feels comfortable having in their home, or that a child feels safe learning from, or that a patient trusts to assist with rehabilitation, the emotional response to its face is not a trivial consideration. It’s the consideration.
The honest answer to “will we cross the uncanny valley?” is probably “yes, eventually, for robots that are sufficiently expensive and specifically designed for contexts where crossing matters.” The Columbia lip-movement research, the LLM conversation studies, and advances in silicone skin and micro-actuator facial expression systems are all pushing the right side of the valley upward. But for the next decade of commercial humanoid robotics, the dominant strategy will almost certainly be avoidance rather than crossing—designing robots that are useful, functional, and friendly without pretending to be something they aren’t. The uncanny valley is a real phenomenon with real psychological mechanisms behind it. The most practical response, for now, is to respect it rather than try to solve it.
We cover the uncanny valley alongside the mechanical engineering, AI integration, and commercial deployment of humanoid robots across 24 lectures in our Humanoid Robots & Drones course—including why the companies spending billions on robot bodies are making very specific choices about what those bodies look like.