§ THE EMBODIED FLY
The built result. A real FlyWire connectome, run as a spiking brain, drives a real fly body: a looming threat becomes a Giant-Fiber spike becomes an escape bolt, and the loop closes through the physics. This is the top of the line the rest of the bench climbs — the place the real circuit finally drives the body.
The brain sits between a real sensor and a real body, and the loop closes through the physical world — not a shortcut. Numbers track left threat (azimuth 90): the Giant Fiber fires 133.33 Hz, the drive is 0.178, and the body bolts right (away). ⓘ hover, tap, or focus any stage for the mechanism — what each one actually computes.
We took the electron-microscope wiring diagram of an entire adult Drosophila brain — 138,639 leaky-integrate-and-fire neurons, ~15 million synapses, the real FlyWire v783 connectome — ran it as a spiking network, and used it to drive a biomechanically real fly body in MuJoCo, in a closed sensory-motor loop. A looming “predator” is seen by the fly's looming-detector neurons; their spikes propagate through the actual wiring to the Giant Fiber escape command neuron; and that neuron's firing makes the simulated fly bolt away — after which its movement changes what it sees, and the loop repeats.
It is a hand-built, miniature recreation of Eon Systems' embodied-fly demo: their integration of the same published parts (the FlyWire connectome, the Shiu LIF brain, the NeuroMechFly body). A research/demonstration platform, not a complete fly. The honest line, kept throughout: this shows the real connectome routing a looming cue to an embodied escape — not a calibrated escape threshold.
§ THE ESCAPE CIRCUIT · WHERE IT SITS
The circuit Vishal kept coming back to. Inside that brain is the real escape sub-circuit — three identified cell types, the cleanest known looming-detector → escape pathway biology hands us. The looming detectors LC4 (≈ angular velocity) and LPLC2 (≈ angular size) sit laterally in the lobula complex, behind each eye, and converge ipsilaterally onto DNp01, the Giant Fiber, which descends toward the body. The wiring is asymmetric, and it is drawn so you can see it: each edge's thickness is its synapse count.
LC4104 neurons · cholinergic
A lobula columnarvisual projection neuron tuned to a looming object's angular velocity— how fast its image is expanding. The “it's coming fast” channel. Sits laterally in the lobula complex, right behind the eye.
LPLC2210 neurons · cholinergic
A lobula-plate/lobula columnar projection neuron tuned to angular size— loom geometry, the object filling the eye near collision. The “it's getting big” channel. Also lateral, in the lobula complex.
DNp01the Giant Fiber · 2 (1/side)
The descending command neuron for fast escape — one per hemisphere, the largest axon in the fly. LC4 and LPLC2 converge on its lateral dendrite; it sums size + velocityand carries the “escape now” command down toward the ventral nerve cord.
Where they come from. Every cell here is a real, addressed neuron in the FlyWire v783 connectome — the electron-microscope wiring diagram of a whole adult Drosophila brain. CX-1's curation pulls 104 LC4 + 210 LPLC2 → 2 DNp01 (316 cells total), and all 316 resolve in the brain we run — no version drift, because both are keyed to v783. The convergence is ipsilateral: each eye's detectors drive that side's Giant Fiber.
Why these three. This is textbook escape wiring, and the most tractable place a connectome can drive a body: a small, self-contained sub-circuit (316 cells, not 138,639) with a known function. von Reyn et al. 2017 showed the Giant Fiber sums angular size and velocity to time a takeoff; Ache et al. 2019 dissected the LC4 (velocity) vs LPLC2 (size) division of labor converging on it. Driving LC4 + LPLC2 in the live connectome makes DNp01 fire, and the asymmetry is real: the right Giant Fiber carries more converging synapses than the left — 431/374 for LC4, 622/458 for LPLC2 — which is exactly why the right Giant Fiber out-fires the left in the live run. The brain outline is a hand-drawn schematic for placement, not a literal neuropil render; the counts and synapses are real — but in the result below, these exact 316 cells get placed where they really sit and lit by the run, right next to the fly.
§ THE GIANT FIBER · AS AN INSTRUMENT
Drive LC4 and LPLC2 in the live connectome and the Giant Fiber fires; silence them and it goes quiet. The curve below is the real brain-only sweep — and the toggle is the proof that the Giant Fiber only fires through these inputs.
And in the body: driven, the Giant Fiber fires, its rate becomes the escape drive, and the trained body bolts away— the left/right escape runs below. The Giant Fiber is the causal link between this curve and the fly's motion.
size + velocity summation · at 150 Hz drive
sub-additive: 109 + 160 = 269 ≫ 184.5. The Giant Fiber sums size and velocity with diminishing returns — it does not just add them.
§ THE RESULT · LOOMING → ESCAPE
The payoff, and the honest version of Eon's “brain lighting up”: the real body and the real circuit at the same instant. Pick a condition and watch one escape closely — the threat enters the fly's world on the left, the escape circuit fires on the right, on one shared clock. Left threat → Giant Fiber 133.33 Hz → bolts right (away); right threat → 100 Hz → turns left (away); baseline → Giant Fiber silent, the fly just walks. The brain is the causal link.
loading the connectome…
loading
Eon's “simultaneous” view, told straight: the real body in its world on the left, the real escape circuit on the right, on one clock. The threat appears, the threat-side LC4/LPLC2 warm to amber, the Giant Fiber blooms gold, and the fly bolts right (away) — same instant, the marks calling threat onset and the escape pivot. Baseline stays dark: no loom, the Giant Fiber is silent, the fly just walks. The cloud is lit by the actual LIF activity we computed (per-window hz_L/R and dnp01_L/R), not predicted glow; the clip is 4 s and the brain trace ~1.2 s of sim time, so they co-progress by playback, not by the millisecond.
316 real FlyWire FAFB-v14.1 positions · only the computed circuit neurons light up · the dim grey volume behind them is a 40k-point full-brain backdrop in the same FlyWire frame — resting brain, positions only, not computed activity. Drag the brain to orbit.
Then all three runs side by side — the numbers behind the clip above. The Giant-Fiber trace, the top-down body path, and the escape signature for left, right, and baseline, on shared axes so the differences read.
bolts right (away)
loading trace…
turns left (away)
loading trace…
walks (Giant Fiber silent)
loading trace…
The same three runs as the paired view above — here the numbers, side by side (the clips live once, up top). They share the seed and warm-up, so they diverge only at looming onset (the orange mark). The Giant-Fiber trace is the brain causal link: no loom → GF silent → the fly merely walks; a loom routes through the real connectome to an escape. The escape signature is the turn relative to the no-threat baseline (the walking gait itself drifts, so net heading is meaningless) — left threat turns negative (bolts right), right threat positive (turns left). Brain (75 windows) and body (301 steps) are plotted on their own t_s, never joined by index.
§ BUILT ONCE, THEN STEPPED
The make-or-break constraint. Building the spiking network reads a ~100 MB connectivity table; rebuilding it every window would take minutes. So the brain is built once per condition, then stepped with a runtime-settable input rate — brain and body advancing in lockstep every 15 ms.
Timings from the WP-D export run (the bundle is reproducible — seeded, re-running the script reproduces it byte-for-byte).
What it costs.“Built once, then stepped” is what makes the loop tractable, but the choices it forces are real drawbacks, not free wins:
- The 15 ms sync may be too coarse.Brain and body advance in one lockstep window; a real escape's decisive dynamics can be faster than that, so the fastest part of the behavior is under-sampled. Fine for a turn-and-flee, likely too slow for, say, a wing-beat-timed takeoff.
- The single-DN readout is quantized and flickers. The escape command is read from one DNp01 per hemisphere over that window, so the mean rate lands in ~33 Hz steps and can swap a step run-to-run — the peak split here (133.33 vs 100 Hz) is on that grid. Averaging more GF-pathway neurons / longer windows would smooth it.
- No sub-window dynamics. Within a window the drive is held constant, so anything happening faster than 15 ms — fine spike timing, rapid sensory change — is invisible to the body. The loop sees a stepped approximation, not the continuous trajectory.
§ HONEST LIMITS
Honesty is the brand — these carry the same weight as the headline. What this is, and what it isn't:
- Not a calibrated threshold.The claim is “the real connectome routes a looming cue to an embodied escape,” not a claim about the escape threshold or stimulus selectivity. In isolation the Giant Fiber saturates easily — real selectivity lives in whole-brain inhibition this stack doesn't yet capture. Don't read the response curve as “the fly escapes at N Hz of looming.”
- Coarse, flickering readout. One DNp01 per hemisphere, read over a 15 ms window → the mean Giant-Fiber rate is quantized in ~33 Hz steps and flickers. The per-side peak split (here 133.33 vs 100 Hz) can swap ±1 step run-to-run; it is seeded here so the committed bundle is reproducible. The direction is bearing-driven and deterministic regardless of seed.
- Hand-tuned, sparse coupling. Three interface mappings (looming→Hz gain 150, GF→drive ref 150 / peak 0.2, 15 ms sync) are knobs, not measurements. One descending neuron stands in for the whole escape command; the looming front-end is a hand-built stand-in for the visual pathway; v1 drives LC4 and LPLC2 identically (splitting size→LPLC2, velocity→LC4 is the faithful next step).
- Inherited body asymmetry, ground-only.The left-strong / right-weak escape is the trained body controller's asymmetry, reported not hidden. It is a ground turn-and-flee — no true takeoff yet.
- A predicted, snapshot cartoon. One connectome; neurotransmitter signs are predicted, not measured; leaky-integrate-and-fire neurons omit dendritic computation, channels, plasticity (no learning — the wiring is frozen), and neuromodulation.
- “Structure → behavior” is a direction explored, not proven. This is the first complete vertical slice of that question, not an answer to it.
§ THE LINE THIS COMPLETES
One controller slot drives one fixed body; the line swaps progressively more biological structure into that slot. Rung 01 is the starting placeholder — a generic NCA null model, the proof a local rule can walk the body, with nothing biological in it. Rung 02 closes the proprioceptive loop. Rung 03 — this page — fills the slot with the real FlyWire connectome. That top rung is the climax; the rungs below it are the run-up, not the goal.
§ REFERENCES
The published parts this is built from — the connectome, the brain model, the body, the escape circuit, and the reference system.
- A leaky integrate-and-fire computational model based on the connectome of the entire adult Drosophila brainThe FlyWire connectome turned into a spiking LIF brain — the model run here.
- Neuronal wiring diagram of an adult brain (FlyWire)The connectome itself — the measured v783 wiring the brain is built from.
- NeuroMechFly v2: simulating embodied sensorimotor control in adult DrosophilaThe MuJoCo fly body (FlyGym) the brain drives.
- Feature integration drives probabilistic behavior in the Drosophila escape systemThe Giant Fiber sums size + velocity; single-spike timing sets the takeoff.
- Neural basis for looming size and velocity encoding in the Drosophila giant fiber escape pathwayLC4 (velocity) + LPLC2 (size) → DNp01 — the circuit reproduced here.
- How the Eon team produced a virtual embodied flyThe reference embodied fly this is a hand-built recreation of — connectome brain + body on MuJoCo, run on CPU, GPU, and neuromorphic backends (incl. Loihi 2).