Neuromorphic Vision · RTL-Verified · Event-Based
LIBELLULA's architecture is drawn from three bodies of research: dragonfly target-interception behaviour, the cellular-level understanding of motion computation in the fly optic lobe, and the small-target motion detector neurons that let a perched dragonfly pick one tiny moving object out of clutter. This page explains all three and where each one lands in the RTL.
Chapter I
Dragonflies are among the most successful aerial predators by capture rate: Olberg and colleagues recorded success above 90 % in the field. Unlike most predators that react to prey location, the dragonfly computes where prey will be and flies to that point — an internal prediction loop running faster than any conventional sensorimotor reflex.
Two populations of neurons carry this. In the lobula, small-target motion detector (STMD) neurons, studied in depth by David O’Carroll and Steven Wiederman, respond to a tiny moving object and to almost nothing else, including the same object’s motion when it is part of a wider pattern. Downstream, target-selective descending neurons (TSDNs), first characterised by Robert Olberg, carry the selected target’s position and direction straight to the flight motor circuits. The dragonfly does not deliberate. The prediction is the action.
Studies of dragonfly interception trajectories show flight paths consistent with predictive pursuit (constant-bearing, decreasing-range geometry) rather than reactive tracking, and Wiederman’s group has shown that STMD neurons modulate their gain ahead of where the target is expected to appear. The circuit fires ahead of the target’s arrival.
Chapter II
In 1956, Bernhard Hassenstein and Werner Reichardt described a model of directional motion detection derived purely from behavioral experiments on walking beetles. The model contains no optics, no biology — only logic: if a stimulus activates two adjacent detectors in sequence, and the delayed output of the first is multiplied with the instantaneous output of the second, the result is direction-selective.
Photoreceptor A Photoreceptor B
│ │
├──── delay (τ) ─────────┤
│ │
└──── × ─────────────────┘ ← multiplication
│
Direction signal
Mirror-symmetric subunit (A←B) subtracted:
└──── × (null direction) ────┘
│
Net directional output
Key operation: temporal correlation via delay + multiply
The model predicted, from behavior alone, that motion computation would require: a temporal delay between adjacent channels, a nonlinear interaction (multiplication), and subtraction of opponent directions. All three were subsequently confirmed in electrophysiology — decades later.
What made the model remarkable was its predictive power across stimulus types. Gratings, apparent motion sequences, stochastic flicker — the Hassenstein-Reichardt correlator reproduced recorded neural responses quantitatively across all of them. It remained the governing framework for insect motion detection for nearly 70 years.
Chapter III
The cellular elements of the Reichardt detector were identified in Drosophila over the last decade, principally through work in Axel Borst's group at the Max Planck Institute of Neurobiology and connectomics efforts at Janelia Research Campus.
The fly optic lobe has a columnar organisation: photoreceptors → lamina → medulla → lobula → lobula plate, each column processing one facet of the compound eye. Within this structure, T4 and T5 neurons are the primary motion-sensing cells.
Downstream of T4/T5, lobula plate interneurons perform the subtraction stage — they receive excitatory input from one directional layer and inhibit the adjacent opposing layer. This implements the null-direction suppression that sharpens the tuning curve beyond what either mechanism achieves alone.
aer_rx — event ingestionlif_tile_tmux — temporal encoding (LIF)delay_lattice_rb — retinotopic ring-buffer delayreichardt_ds — preferred-direction enhancementburst_gate + conf_gate — confidence gatingab_predictor — α–β forward trajectory estimateChapter IV
For decades, two competing theories of directional selectivity existed in parallel. Hassenstein and Reichardt proposed that motion detection works through preferred-direction enhancement: the delayed signal from one photoreceptor is multiplied with the undelayed signal from the next, amplifying responses when motion crosses in the preferred direction.
Barlow and Levick proposed instead null-direction suppression: a delayed inhibitory signal from the null direction vetoes the response before it can propagate.
Both theories were right. Research using apparent-motion stimuli at individual facets of the fly eye showed that the answer depends on where within the receptive field you probe: the leading edge shows preferred-direction enhancement; the trailing edge shows null-direction suppression; the center shows both simultaneously.
| Mechanism | Location in receptive field | Effect on tuning curve | LIBELLULA module |
|---|---|---|---|
| Preferred-direction enhancement | Leading edge | Amplifies preferred-direction response | reichardt_ds multiply path |
| Null-direction suppression | Trailing edge | Reduces null-direction response | reichardt_ds inhibit path |
| Combined (dual mechanism) | Full receptive field | Tighter tuning than either alone | Full reichardt_ds module |
The biophysical substrate of this dual mechanism was identified through receptor mapping on T4 cell dendrites. GluCl-α receptors (glutamate-gated chloride channels — inhibitory) are concentrated at the dendritic tips, where Mi9 input arrives. Nicotinic acetylcholine receptors cluster in the center, where Mi1 input arrives. GABA-A receptors sit at the base, driven by Mi4.
When a bright edge crosses the receptive field in the preferred direction: Mi9 (an OFF cell) is silenced — its chloride channel closes — causing the dendritic tip's input resistance to rise, which amplifies the central excitatory Mi1 signal. Motion in the null direction causes the opposite sequence, suppressing the response. The geometry of the dendrite performs the computation. No learning. No weights. Fixed physics.
Chapter V
Wide-field motion detection answers “which way is the world moving?” A hunting dragonfly needs a different answer: “is there one small thing moving on its own?” STMD neurons give it. They respond to a target a degree or two across, moving against a textured background, and stay silent for the background’s own motion.
In 2008 Wiederman, O’Carroll and Shoemaker proposed a model that produces this selectivity without any template for the target’s size. Adapt each photoreceptor so that sustained contrast fades. Split the signal into ON and OFF channels. Inhibit each channel with its spatial surround, so that extended edges cancel and small ones survive. Then delay one channel and correlate it with the other at the same location: a dark object passing a point produces an OFF transient followed, when it leaves, by an ON transient, and only that ordered pair passes. They called it the elementary STMD, and it reproduced the recorded tuning of the neurons for size, velocity and contrast.
Event cameras produce the model’s input directly. A DVS pixel emits an OFF event when it darkens and an ON event when it brightens, and nothing while the scene is still. LIBELLULA’s detector is a per-event implementation of the 2008 chain: a per-pixel activity filter and flicker habituation in place of photoreceptor adaptation; the OFF-then-ON correlation at each pixel; centre-surround inhibition on the accumulated response; a small pool of tracks that keeps one target over time, as the TSDN carries one target to the wings; and two adaptive stages the recordings forced, for sun glare and sensor noise.
stmd_frontend — activity filter, flicker habituation, glare and noise adaptationstmd_backend — opposite polarity at the same pixel within 3–100 msstmd_backend — 3×3 centre against a 9×9 ring of response and raw activitystmd_backend — eight-track pool, α–β update, confirmation, stationarityOne result is worth stating here because it bears on the biology rather than the engineering. On the public EV-UAV recordings (tiny drones against sky, a static DAVIS346 camera), the centre-surround inhibition and the tracker carry the detector’s performance; the OFF-then-ON correlation, the stage the model is named for, contributes under one percent in every configuration tried. Whether that reflects the data (a drone against sky is not a dark target in clutter) or the implementation is an open question. The stage is kept, switchable, and reported as it is.
Chapter VI
Dynamic Vision Sensors (DVS cameras) output asynchronous spike events — one per pixel, per polarity change, timestamped to the microsecond. This is functionally equivalent to the output of the fly’s photoreceptor and lamina stage: sparse, temporally precise, polarity-coded signals rather than continuous luminance values.
The biological circuits that evolved to process exactly this type of input, and to extract direction, a selected target and its future position from it within a few milliseconds of neural time, are the Hassenstein-Reichardt pipeline and the STMD chain. LIBELLULA maps both onto synthesizable RTL, preserving the essential structure: spatial delay, temporal correlation, direction-selective multiplication, surround inhibition, confidence gating, and predictive integration. In logic the same operations take nanoseconds; the structure, not the speed, is what is borrowed.
The result is deterministic behavior at every stage. The output of each module is traceable from input event to predicted coordinate or target label without any learned parameters, stochastic inference, or black-box processing. This is what these biological circuits do — and it is the design target for LIBELLULA.
Sources & further reading
Synthesizable Verilog-2001 RTL: the motion-prediction core, 42 benches, and a small-target detector verified bit for bit against a fixed-point C reference.