LIBELLULA v11

In 1956, Bernhard Hassenstein and Werner Reichardt performed a landmark experiment on the beetle Chlorophanus viridis. By analysing the beetle's optomotor turning response to controlled moving patterns, they deduced the minimal computational structure required to detect the direction of motion: two adjacent photoreceptors, a temporal delay applied to one signal, and a correlation (multiplication) of the delayed signal with the instantaneous signal from the neighbouring receptor. This structure is now called the Elementary Motion Detector, or Reichardt detector. (Hassenstein & Reichardt, 1956, Z. Naturforsch.)

This is not merely historical curiosity. Reichardt detectors, in various elaborated forms, have been confirmed in the optic lobes of flies, dragonflies, locusts, and bees. In dipteran flies, the T4 and T5 neurons of the medulla implement exactly this correlation architecture for ON and OFF contrast pathways respectively. (Borst & Euler, 2011, Neuron; Maisak et al., 2013, Nature). In dragonflies, the same principle operates in the medulla-to-lobula pathway, feeding directional motion signals into the small target motion detector (STMD) neurons downstream.

LIBELLULA's delay_lattice_rb and reichardt_ds modules are a direct silicon translation of this architecture. The delay lattice provides configurable temporal delays across four cardinal directions using ring buffers—an engineering implementation of the graded synaptic and dendritic delays found in biological neural arbors. The Reichardt detector module then performs the correlation step with leaky integration, converting raw temporal gradients into signed directional motion vectors.

Why the Reichardt detector rather than gradient-based optical flow or a learned motion estimator? Because it computes direction selectivity in fixed, feedforward logic with minimal latency. There is no iterative solver, no frame differencing, and no weight matrix. A single pass through the delay-correlate structure yields a motion vector—the same economy that lets a dragonfly's optic lobe process motion in under a millisecond with a few hundred neurons.

Dragonflies do not chase prey. They intercept it. High-speed videography of hunting dragonflies shows that they launch along a trajectory that will intersect the prey's future position, not its current one. This requires forward prediction: estimating where the target will be in tens of milliseconds, and steering toward that point rather than the visible target. (Olberg et al., 2000, J. Comp. Physiol.; Mischiati et al., 2015, Nature)

The geometry of proportional navigation—steering to maintain a constant bearing angle to the target—is the optimal solution for this interception problem, and dragonflies approximate it closely. But proportional navigation requires velocity estimation and trajectory extrapolation, not just position sensing.

The neural substrate for this is found in the target-selective descending neurons (TSDNs), which encode not just the current position of a small target but its projected trajectory relative to the dragonfly's own motion. These neurons integrate directional motion signals from upstream STMD neurons—themselves fed by the Reichardt-detector arrays described above—and produce motor commands that anticipate the target's future position.

LIBELLULA's α–β predictor is a clean abstraction of this behaviour. It does not classify targets or estimate intent. It takes the directional motion vectors from the Reichardt stage, applies a smoothing-and-extrapolation filter (the α–β tracker), and outputs a predicted position: p + v·Δt. The prediction horizon Δt is configurable and selectable at runtime, allowing the same hardware to serve applications from microsecond-scale reflex avoidance to tens-of-milliseconds trajectory planning.

A dragonfly hunting in a swarm does not track every mosquito. It locks onto one target—often the nearest or most salient—and ignores the rest until the chase is complete. This is winner-take-all selective attention, implemented not through learned weighting but through lateral inhibition and gain modulation in the CSTMD1 (Centrifugal Small Target Motion Detector 1) neurons. (Wiederman & O'Carroll, 2013, Curr. Biol.; Wiederman et al., 2017, eLife)

CSTMD1 responds preferentially to small, high-contrast targets moving against a cluttered background—exactly the signature of prey. When multiple targets are present, the neuron's response to the attended target is enhanced while responses to distractors are suppressed. This attentional lock persists even when the attended target is briefly occluded, suggesting a short-term predictive memory.

LIBELLULA's burst gate and confidence gate implement the same principle in deterministic logic. Activity is amplified where motion coherence emerges—where events cluster in space and time, consistent with a real moving object—and damped elsewhere. The burst gate acts as a density filter, requiring a threshold of coincident events before passing a signal downstream. The confidence gate scores the output by combining event rate with directional magnitude, producing a reliability measure that gates the predictor. Together, they enforce hard attentional sparsity: only coherent, salient motion survives to drive prediction.

Insect vision is fast not because neurons fire quickly—biological spike rates are modest—but because the pathway is short. A dragonfly's visual processing from photoreceptor to descending motor neuron involves perhaps 5–8 synaptic stages, each adding a few milliseconds of delay. The entire loop from photon to wing-muscle command can complete in under 30 milliseconds. There is no deep hierarchy, no iterative refinement, no recurrent settling. The architecture is feedforward and shallow, optimised for latency over representational richness.

LIBELLULA's sub-microsecond pipeline is aligned with the same constraint: action must occur before deliberation becomes possible. Six clock cycles at 200 MHz, roughly 30 nanoseconds, is all that separates an incoming event from a predicted output position. There is no opportunity for batching, reordering, or probabilistic hesitation. Every design choice privileges immediacy—from the purely feedforward topology to the fixed-function Reichardt correlators to the single-pass α–β extrapolation.

Dragonfly vision does not learn online in the way mammalian cortex does. A dragonfly hatching from its larval stage can hunt prey within minutes; its predictive interception circuitry is genetically specified and calibrated by development, not by experience. The Reichardt detector parameters, the STMD tuning curves, the TSDN directional preferences—all are largely hardwired. What plasticity exists is slow and structural (seasonal changes in receptor sensitivity, for example) rather than rapid and synaptic.

LIBELLULA, in its current form, preserves this fixed-reflex character. There is no embedded learning loop, no adaptive weight drift, no internal model revision. The delay lattice depths, the Reichardt correlator gains, the burst-gate thresholds, and the α–β filter constants are all set at synthesis time. That absence of learning is itself a biological trace—and a deliberate engineering advantage. A fixed pipeline is formally verifiable, deterministic, and certifiable in safety-critical contexts. Learning is deferred to downstream systems or future co-development, not embedded in the predictive core.