How Praying Mantises See in 3D: The Strange Stereo Vision of an Insect Brain

For most of the 20th century, stereoscopic depth perception was thought to require a large brain. The textbook account was that humans, primates, cats, owls, and a few other vertebrates with forward-facing eyes computed depth from the small differences between left and right eye views, and that this computation was expensive enough to require substantial cortical real estate. Insects, with their small brains and (usually) lateral eyes, were assumed to navigate via parallax and other monocular cues without true binocular stereo.

The assumption was wrong. In 1983, Samuel Rossel published the first demonstration that praying mantises use stereoscopic vision to compute strike distance, a finding that was strange enough at the time to be controversial. Subsequent work has confirmed that mantis stereo vision is real, is used for predation, and operates by a mechanism that is genuinely different from vertebrate stereo vision rather than a miniaturized version of it.

The basic puzzle

A praying mantis catches prey using a ballistic strike of the forelegs. The strike has to land at exactly the right distance: too short and the prey escapes, too long and the mantis cannot grasp before the prey reacts. The strike happens in under 100 milliseconds, faster than the mantis can correct mid-motion, so the distance has to be computed before the strike begins. The mantis is a sit-and-wait predator with a small brain (under a million neurons), and yet it gets the strike distance right essentially every time.

The pre-1983 explanation was that mantises used some combination of image size (prey of known size occupies more visual field when closer), motion parallax (head movements produce image displacements that scale with distance), and accommodation cues (insect eyes do not focus much, so this was a weak hypothesis). Stereo vision was not considered because the mantis brain was thought to be too small to do the computation.

The 1983 demonstration

Rossel's 1983 paper in Nature reported the first decisive experiment. He fitted mantises with tiny prisms over one or both eyes, displacing the apparent position of prey relative to its actual position. If the mantis was using stereo, the displaced views should cause it to misjudge distance in a predictable way (the strike should be aimed where the stereo computation said the prey was, not where it actually was). If the mantis was using monocular cues, the prism displacement should not affect strike distance because the cues are independent of the difference between eyes.

The result was unambiguous: prism-fitted mantises misjudged strike distance in the direction predicted by stereo vision, with the magnitude of the error matching the magnitude of the imposed disparity. The same prisms on one eye only produced no effect, confirming that the cue was binocular rather than monocular. The 1983 paper closed the basic question: mantises use stereo.

The mechanism turns out to be different

The follow-up work over the next thirty years has shown that the mantis is not just doing a small version of vertebrate stereo. The mechanism is different in ways that have taken decades to characterize and that imply that stereo vision evolved independently in this insect lineage and arrived at a different solution.

The decisive paper was Vivek Nityananda's 2018 Current Biology study at Newcastle University. Nityananda and colleagues fitted mantises with miniature 3D glasses (using teal and amber filters analogous to anaglyph 3D glasses for humans) and showed mantises 3D movies of prey. The mantises struck at the simulated prey positions, confirming the stereo mechanism in a controlled way. The more important finding was that mantises responded to moving prey in the stereo movies but did not respond to static prey, even when the static prey was clearly visible at a stereo-defined distance.

The contrast with human stereo vision is sharp. Human stereo computes depth for both moving and static objects, using the absolute geometric disparity between left and right retinal images. Mantis stereo seems to compute something more like "where is the moving thing relative to me," using changes in disparity (motion-defined depth) rather than absolute disparity. The two systems solve the same predation-relevant problem (catch the moving thing at the right distance) using mathematically different algorithms.

The neuroanatomy

The mantis brain has, by vertebrate standards, almost no real estate available for stereo computation. The total brain has under a million neurons, the optic lobes account for perhaps half of those, and only a small fraction of optic lobe neurons receive binocular input. The computation seems to be happening in a relatively small population of neurons in the lobula complex, with binocular comparison happening at an earlier processing stage than in vertebrate systems.

The Karin Nordstrom and Trevor Wardill labs have identified what appear to be the mantis equivalent of disparity-tuned neurons, responding selectively to moving features at specific binocular disparities. The functional architecture is recognizably parallel to vertebrate disparity-tuned cells in V1 and beyond, but the cell count is vastly smaller and the response properties are tuned for motion rather than for general depth.

The strange detail is that the mantis stereo system does not seem to be tuned for fine depth discrimination at large distances. The behavioral data suggests reasonable depth accuracy out to about 2-3 cm (the strike range) and much poorer accuracy beyond that. This is a sensible engineering choice: the mantis only needs depth information for prey within strike range, so dedicating limited neural resources to fine depth at long range would be wasteful.

Convergent evolution

The mantis is not the only insect with stereo vision, though it was the first one discovered and remains the best-characterized. Subsequent work has identified stereo or stereo-like depth processing in dragonflies (which use it for prey interception during flight), some robber flies (also predatory), and possibly some jumping spiders (though spiders are not insects and use a different visual organization with multiple specialized eye pairs).

The pattern is consistent: predatory arthropods with forward-facing eyes that need accurate distance to a moving target tend to evolve some form of stereo computation, using neural machinery much smaller than vertebrate stereo systems. The independence of these inventions (mantises, dragonflies, and robber flies diverged 300+ million years ago) suggests that the basic computation is not as expensive as vertebrate biology had implied.

The other side of the comparison is the absence of stereo in most insects. Bees, butterflies, beetles, ants, and most other insects either do not have stereo vision or use a much weaker form of it. The selection pressure that produces stereo seems to be specifically the combination of predation, forward-facing eyes (which most insects do not have), and a fast strike that has to be aimed before it begins.

The applied research surface

The mantis has become a model organism for neuroethology of vision, with active programs at Newcastle, Cambridge, Tubingen, Konstanz, and several other labs. The research is partly fundamental (how does stereo work in a small nervous system?) and partly applied (can the small-circuit solution be useful for robotic vision?).

The robotics application is real but underdeveloped. Modern depth-sensing for robots typically uses either active sensors (LIDAR, structured light) or stereo cameras with substantial computational backing. The mantis-style approach of motion-defined stereo with minimal computation has obvious appeal for power-constrained applications (insect-scale drones, low-power wearable sensors), and Cambridge groups have demonstrated small-form-factor stereo cameras using mantis-inspired algorithms that run at much lower power than conventional stereo. The application surface is still emerging, but the proof of concept is there.

The neuroscience implications are also interesting. The mantis stereo circuit is small enough to potentially be fully characterized at the cell-and-connection level, in a way that would be impossible for vertebrate stereo systems. The connectomic work on Drosophila optic lobe is the obvious analog, though mantis brains are not the focus of the major connectomics projects. A full circuit description of mantis stereo would be one of the cleanest cases of "depth computation as algorithm" available.

Three observations

First, the assumption that complex perception requires large brains was based on a small sample of vertebrate examples and turned out to be wrong in at least one important case. The pattern across recent comparative neuroethology is that this happens repeatedly: capabilities thought to require mammalian cortex turn up in birds, then in fish, then in arthropods, then in cnidarians, with the threshold for "complex" steadily lowering as smaller nervous systems are studied more carefully.

Second, the mantis case is a clean example of independent reinvention producing functionally similar but mechanistically different solutions. The vertebrate and mantis stereo systems are not homologous; they evolved independently and arrived at different algorithms tuned to different selection pressures. The deeper point is that "stereo vision" is not a single thing; it is a problem with multiple solutions.

Third, the discovery timeline (basic phenomenon in 1983, mechanism details still being filled in in the 2020s) is typical of comparative neuroscience. The questions that are easy to ask (does this animal do X?) get answered first; the questions that are hard to ask (how exactly does the circuit compute X?) take decades and depend on instruments and techniques that develop in parallel with the questions.

The broader pattern is that the inventory of perceptual capabilities in the animal kingdom is consistently richer than vertebrate-centered biology textbooks suggest. The mantis joins the cuttlefish color vision puzzle and the bee polarization compass and the bat echolocation system as cases where a small nervous system implements a sophisticated perceptual computation in a way that does not fit the canonical mammalian template. The textbook generalization that complex perception requires big brains is true on average but the exceptions are numerous and turn up exactly where selection pressure is strong.