The uncanny valley, explained

In 1970, the Japanese roboticist Masahiro Mori published a short essay in an obscure journal, Energy, with a hand-drawn graph that turned out to be the most important chart in this industry's history. His observation: as a machine looks more human, our affinity for it rises — until it gets close, at which point affinity collapses into revulsion before recovering as the likeness approaches perfection. He called the dip bukimi no tani: the uncanny valley.

Every machine in our directory lives somewhere on that curve, and every design decision its maker took is a bet about how to survive it.

Why the valley exists

Half a century later there's still no settled answer, but the leading explanations rhyme: a near-human face triggers the brain's specialized face machinery, which then detects mismatches — eyes that don't quite converge, skin that doesn't micro-move, a smile that starts a beat late. Some researchers frame it as category confusion (human or thing?), others as a hair-trigger for illness or death cues. What matters practically is the consensus finding: the effect is real, measurable, and strongest when most signals say human and a few say otherwise.

Mori's second insight is the one that matters

The famous graph has a second curve everyone forgets: movement amplifies everything. A still corpse is unsettling; a moving one is horrifying. The same multiplier hits robots — which is why press photos systematically flatter these machines and why we embed video on every profile. A face that passes in a photograph must then blink at the right rhythm, track your eyes, and land its lip-sync within tens of milliseconds; each miss digs the valley deeper. This is exactly why our methodology scores motion and voice separately from the face itself, and why unscripted footage outweighs any launch video.

The three strategies

Watch the field through Mori's graph and every maker is running one of three plays.

Stop before the valley. Most humanoids deliberately stay machines: NEO wears a friendly knit suit, Apollo communicates through displays, Digit has an LED head. Warm, legible, zero uncanny risk — and zero chance of being mistaken for company.

Stand at the edge. Engineered Arts gave Ameca the world's most expressive robot face and then painted it grey — brilliant expressions, explicitly non-human skin. The mismatch detector never fires because the face never claims to be human. It's the cleverest dodge in robotics, and the reason Ameca's videos charm rather than unsettle.

Charge through. Aria, Grace, the UBTech U1 — silicone skin, human hair, the full claim. This is the companion industry's necessary gamble: presence is the product, so the valley must be crossed, not avoided. The known results are instructive. Sophia's face — remarkable in 2016 — reads stiff in motion today. Aria holds up best in stillness and conversation, helped by a wheeled base that never asks her body to move convincingly. The U1 raises the stakes by walking; whether a full-body gait deepens the valley or crosses it is precisely what September's deliveries will show.

Does the valley ever close?

Two honest observations. Exposure seems to matter — a generation raised on CGI faces and video-call filters reports less discomfort with near-human machines than Mori's contemporaries did, and the valley may be partly cultural, shifting as the technology normalizes. And the far side is real: humans are proof that a perfectly human-reading face is possible; it's an engineering distance, not a law of nature. The distance is just much longer than any launch video implies — measured in micro-expressions, latency and gaze, the exact places our realism scores dock the hardest points.

Mori's own advice to designers, fifty years ago, was to aim for the first peak — build something likeable and honestly artificial rather than gamble on near-human. The whole companion industry is now a wager that he was wrong about where to stop. The next two years of owner reports will grade it.