How Archerfish Spit: The Strange Multi-Media Physics of Aerial Predation

Toxotes jaculatrix, the banded archerfish, is a small Indo-Pacific freshwater fish (10-15 cm body length) that hunts terrestrial insects above the water surface by shooting them down with jets of water from below. The fish lines up under the prey, ejects a focused stream of water from its mouth, knocks the insect off its perch, and catches it as it falls to the water surface. The shooting accuracy at 1-2 meters from the water surface is several centimeters per shot, with capture rates of 40-50 percent on first attempts and substantially higher across multiple shots.

The behavior was first described in detail by Hugh Smith in 1936 (Natural History magazine, the predecessor to Hugh Smith's more famous descriptions of fish hunting strategies). Schuster lab at Bayreuth has spent two decades quantitatively characterizing the mechanics and the learning behind them, producing a substantially revised account of what the textbooks described.

Three independent physics problems

The shooting requires solving three physics problems that interact: refraction at the air-water boundary distorts the apparent position of the aerial prey relative to its actual position; ballistic trajectory through air means the water stream falls under gravity over the time it takes to travel to the target; and jet compression during flight (later water catching up to earlier water due to temporal velocity modulation in the mouth) amplifies the impact momentum at the target.

The refraction correction is the most counterintuitive. An aerial object at 1 meter above the water surface, viewed from below at a 30-degree angle from vertical, appears to be at a different position than its actual location because light refracts at the interface. The fish has to aim at the actual position, not the apparent position. The Schuster group's playback experiments have established that the fish does not learn refraction correction by trial and error in adulthood; the calibration appears to be acquired during juvenile development with some genetic predisposition.

The ballistic correction is the second. The water stream is a projectile in air after it leaves the mouth, subject to gravity and air resistance. Over 1-2 meter distances at the velocities involved (3-5 m/s exit velocity), the drop is modest but non-zero, and the shot has to aim above the target to land on it. The drop varies with distance, so the aim has to adjust based on distance, which the fish has to estimate from the underwater view alone.

The jet compression is the third and was characterized only recently. Vailati and colleagues (Journal of Fluid Mechanics 2012, Italy) showed that the water stream the fish produces is not a uniform-velocity jet but a temporally modulated stream where later water exits the mouth faster than earlier water. The result is that the slower leading edge gets caught by the faster trailing edge during flight, compressing the jet into a higher-momentum impact at the target. The compression amplifies the effective kinetic energy delivered to the prey by a factor that depends on distance and modulation profile.

The Schuster lab corrective

The mid-20th-century textbook account of archerfish shooting was that the behavior was a fixed action pattern: the fish has a hardwired program for shooting that operates the same way for every shot, with limited adjustment to circumstances. This account was consistent with the small fish brain (approximately 5 million neurons in the relevant regions, against the 86 billion in a human brain) and with the speed of the behavior (decision-to-shot timing of approximately 100 milliseconds).

The Schuster lab work since the early 2000s has demonstrated that this account is substantially wrong. The behavior is learned, requires practice for accuracy, supports observational learning from other archerfish, and adjusts in real time to circumstances. Naive juvenile archerfish are inaccurate shots; their accuracy improves over weeks of practice with measurable learning curves. Experienced archerfish observing other archerfish shoot improve faster than archerfish that do not have the observational opportunity.

The observational learning result (Schuster lab, Current Biology 2006) was particularly striking because observational learning had been considered the province of larger-brained mammals and birds. Demonstrating it in a small-brained fish required careful experimental design (controlling for tank conditions, prey types, individual variation) but the result was robust across multiple experimental setups.

The motion prediction component

The hunting performance includes a motion prediction component that is even more striking than the refraction correction. Archerfish do not just shoot stationary targets; they also shoot moving targets, where the shot has to aim where the prey will be by the time the water arrives, not where the prey is when the shot is fired.

Davis and Smith (2015 work on Toxotes) showed that archerfish predict prey trajectory in the pre-shot milliseconds and aim the shot at the predicted future position. The prediction accuracy is high enough that moving targets are captured at rates comparable to stationary targets, which would not be the case for pure reactive shooting.

The motion prediction is implemented in a dedicated visuomotor circuit that has been partially characterized through neuroanatomy. The relevant brain regions appear to include parts of the optic tectum and the cerebellum, with substantial connectivity to motor control areas. The architecture is recognizable as a small-brained version of the predictive interception circuits characterized in dragonflies (Mischiati et al, Nature 2015) and in primate-like motor control.

The cooperative dimension

Schuster lab work has also characterized cooperative shooting in some archerfish species, where multiple individuals coordinate shots at the same prey, with one fish doing the first knockdown and another being positioned to catch the falling prey. The coordination is not strict (any nearby archerfish will grab a falling prey) but the positioning during shot setup suggests some level of social anticipation.

The cooperative dimension is documented mostly in field observations and is harder to characterize quantitatively than the single-fish shooting. It suggests that the cognitive architecture supporting shooting includes some representation of conspecifics' actions, which would be consistent with the observational learning capacity but goes a step further toward social cognition.

The applied interest

The applied interest in archerfish shooting is in firefighting and water-cannon design. The biological mechanism produces high-momentum focused water delivery from a small reservoir at modest exit velocity, which is exactly the optimization target for handheld firefighting equipment and some industrial cleaning applications.

The biomimetic translation has been slow. The fluid dynamics of the jet compression depend on subtle temporal modulation of the exit velocity that is easy to characterize but harder to replicate in an engineered jet. The biological mechanism uses muscle-controlled mouth geometry to produce the modulation; engineered equivalents need pump-and-valve control systems that introduce latency and complexity. Modest commercial progress as of the mid-2020s.

Three observations

First, the textbook account of small-fish cognition was substantially wrong, and sustained experimental attention from a single lab over two decades revised it. The pattern is consistent with comparative biology more broadly: small-brained species turn out to have cognitive capabilities that match the demands of their ecological niches, and the apparent gap with large-brained species is partly an artifact of how much research attention has been devoted to characterizing the small-brained species.

Second, the integration of multiple physics problems (refraction, ballistics, jet compression) into a single behavioral output that operates in 100 milliseconds requires implementation through dedicated circuits, not through general-purpose computation. The 5-million-neuron archerfish brain does not have spare capacity for general inference; it has specialized circuits that solve specific problems quickly. The pattern is consistent with insect cognition (dragonfly predictive interception, mantis stereo vision, honeybee waggle dance) and with songbird vocal learning circuits.

Third, the biomimetic engineering challenge is consistent across these cases: replicating the biological mechanism in engineered form requires not just understanding the principle but also matching the subtle temporal and spatial control that biological tissues achieve. The principle is often clear; the engineering translation is harder than the principle suggests.

The deeper observation is that the inventory of cognitive capabilities documented in vertebrate species has expanded substantially over the past several decades, driven not by broad surveys of many species but by sustained research programs on small numbers of focal species (Schuster's Toxotes, Catania's star-nosed moles, Mischiati's dragonflies, Caldwell-Janik's dolphins). The depth of understanding that emerges from sustained attention is qualitatively different from the breadth of understanding that emerges from broad surveys; the field's current view of vertebrate cognition is shaped disproportionately by a small number of focal species and a small number of long-running programs.

Read more essays and technical writing at anethoth.com — a notebook on databases, distributed systems, biology, and the engineering that holds the world together.