fly-loop

A closed-loop simulated Drosophila: a connectome-derived spiking brain driving a biomechanical body, doing odour-driven walking.

Brain — FlyWire v783 leaky integrate-and-fire whole-brain model (Shiu et al., Nature 634:210–219, 2024), via eonsystemspbc/fly-brain. Body — FlyGym / NeuroMechFly v2, pinned to 1.2.1. Bridge — written here. Nobody has released one.

Target hardware: GTX 1080 Ti (Pascal, sm_61, 11 GB) under WSL2 / Ubuntu 24.04.

---

Read this first

The seam: the connectome stops at the neck

FlyWire covers the brain only. The leg motor neurons live in the ventral nerve cord, which this connectome does not include. Descending neurons are therefore the output boundary of the connectome-derived part of this system.

Everything below them — the central pattern generator, leg coordination, joint trajectories — is hand-written, ships with FlyGym, and is neither connectome-derived nor learned.

odour (FlyGym physics)
  -> Poisson drive to 2,248 ORNs        [connectome-derived]
  -> 138,639-neuron LIF brain           [connectome-derived]
  -> DNa02 left/right firing rates      [connectome-derived]   <-- THE SEAM
  -> left/right rate asymmetry
  -> steering bias                      [HAND-TUNED, see docs/bridge-parameters.md]
  -> FlyGym CPG walking controller      [hand-written, ships with FlyGym]
  -> leg joint torques -> physics

Calling this an end-to-end connectome-driven fly would be false. Roughly the top half is connectome-derived; the bottom half is a conventional robotics controller.

The result, stated plainly

The loop runs, and the fly does steer toward the odour — but the honest headline is that the connectome's odour-to-steering signal is very weak, and one constant I chose is doing a lot of work.

Measured facts, all reproducible from this repo:

1. Unilateral ORN drive does lateralise DNa02, ipsiversively — matching Rayshubskiy et al. Separation between left-drive and right-drive conditions is +14.5 Hz (3.7σ). DNa01's separation is 1.5σ, not usable.
2. But DNa02 carries a large fixed left-bias pedestal unrelated to stimulus side: driving the right ORNs still leaves the left DNa02 at 52 Hz while the right sits near zero. Calibration puts the pedestal at +57.6 Hz against a usable slope of only +8.6 Hz per unit contrast (R² = 0.73). The connectome review predicted this from LAL051, a glutamatergic DNa02 partner with 70 vs 130 left/right synapses.
3. Natural odour contrast is far too small to use. The fly's antennae are ~0.3 mm apart; measured inter-antennal contrast is ±1.8%, which the connectome would convert to ~0.26 Hz of DN differential against a 8–17 Hz Poisson noise floor. The bridge amplifies contrast ~55× to get a usable signal. The transduction is real; the stimulus is not natural.
4. A single DNa02 pair cannot carry a clean steering signal on behavioural timescales. Rate SD is √(r/T): ~12 Hz even at a 400 ms readout window, versus an 8.6 Hz signal. SNR stays of order 1, and the trajectories show it — a weak directional bias on a meandering walk, not clean taxis.

Everything I guessed is enumerated in docs/bridge-parameters.md. The most important entry: at contrast_gain = 12 the odour-dependence test came back negative; at 55 it came back positive. That history is recorded there rather than quietly overwritten.

The statistical bottom line

At n=6 per cell, the intact fly's mirror difference is +11.02 mm, 95% CI [+0.36, +21.68], p = 0.044 — significant, but only just, with a CI whose lower bound is nearly zero. Crossing the sensory wiring removes 96% of it (p = 0.95). Silencing DNa02 removes all of it.

So: a real but marginal odour-driven steering effect, carried by DNa02, on a 55×-amplified gradient. Treat it as preliminary. It is not a robust result and this README should not be read as claiming one.

Other honest caveats

• Symmetric drive does not walk straight. FlyGym's [1.0, 1.0] drifts +10.3°. There is a baseline turning bias independent of anything we do.
• The model has no valence machinery. An odour's sign is imposed by us.
• Absolute firing rates are not comparable to electrophysiology — the Shiu model assumes zero basal firing. Only relative left-right differences are meaningful.
• Shiu et al.'s published model was built on v630; this data is v783. Root IDs are not portable (only ~76.6% overlap), which is why neuron identity is resolved from a checksummed v783 dump rather than hardcoded.
• The DNa01/DNa02 literature is two papers from the same lab (Wilson), using different methods. Not fully independent replication.

---

Results by milestone

M0	WSL2 + CUDA 12.6 + torch cu126, GPU execution verified on Pascal	done
M1	Brain on GPU; sugar→proboscis reproduced; benchmarked	done
M2	FlyGym walking, odour readout, turning modulation, mp4	done
M3	Auditable ORN / DNa01 / DNa02 root IDs from a pinned v783 dump	done
M4	Closed loop, decoupled rates, side-by-side video + DN raster	done
M5	Ablation: silence DNa02, test whether odour-dependence dies	done
M6	Optimisation pass + persistent multi-source world	done

M0 — the Pascal trap is real, but not where it is usually claimed

sm_61 is not in any official PyTorch wheel and never has been. Both torch 2.6.0+cu126 and 2.14.0+cu126 report ['sm_50','sm_60','sm_70','sm_75','sm_80','sm_86','sm_90']. An assert "sm_61" in get_arch_list() gate can never pass.

It works anyway because CUDA guarantees cubin compatibility across one minor revision: the sm_60 cubin runs natively on sm_61. Verified with nvcc, PTX stripped so no JIT could mask it. Pin cu126 anyway — but because cu128+ and CUDA 13.x drop the sm_60 cubin that is carrying us, leaving only compute_90 PTX, and PTX JITs forward only, never backward. Details and the measured compatibility matrix: docs/hardware-notes.md.

M1 — the paper's result reproduces

21 sugar GRNs @ 200 Hz, 1 s:

	measured	upstream v783 band
total spikes	17,393	16,353–17,429	✅
active neurons	403	380–416	✅
MN9 contralateral	100 Hz	must fire	✅
MN9 other CB0701	63 Hz	—	✅

The contralateral MN9 fires harder than the other — the asymmetry Shiu et al. report for unilateral sugar GRN activation.

Throughput: 115 ms of simulated brain time per second of wall clock at batch=1 (0.115× realtime, 0.87 ms/step, 1.45 GB VRAM).

Batching is counterproductive on this GPU — batch=4 takes 8× longer in absolute wall clock than batch=1 for the same simulated duration, because matmul(dense, sparse_csr.T) has no efficient batched path. Aggregate throughput falls from 115 to 58 sim-ms/s. Upstream benchmarks to batch=32; that does not transfer to Pascal.

M4 — why the full brain, not a subset

Measured, not assumed. At 115 sim-ms/wall-s and 1.45/11 GB VRAM, a 3 s episode costs ~26 s of brain compute. A hop≤3 ORN→DN subset (~42,400 neurons) would buy maybe 3× while discarding ~11% of DNa02's input synapses and introducing a second model needing its own validation. The measured numbers say we do not need that trade. Closed loop runs at 0.013× realtime overall — the brain is ~9× faster than the full loop, so physics and per-exchange synchronisation dominate.

Rates are decoupled: physics and brain each advance at their own native 1e-4 s step and exchange information at 200 Hz (every 5 ms = 50 native steps each).

M5 — the ablation passes, but read the caveat

Metric is the mirror difference: mean net_y(source left) − mean net_y(source right). An odour-driven fly goes left for a left source and right for a right source, so this is large and positive. A fly driven by the bridge and the fixed pedestal does the same thing regardless, so it is ~0. Three seeds per cell, 2 s episodes, calibration held fixed across conditions (re-calibrating per condition would silently re-centre the silenced fly and hide the effect).

condition	mirror delta	vs intact
intact	+14.15 mm (effect 1.61)	baseline
DNa02 silenced	+0.00 mm	0%
DNa01 silenced (specificity control)	+22.54 mm	159%

Silencing DNa02 abolishes odour-dependence; silencing DNa01 — which M4a showed is not usable as a readout (1.5σ) — does not. By the stated criterion, it passes.

The caveat you should not skip. The silenced trajectories are [-4.57, -4.55, -4.69] for both source positions — identical to two decimals. That is not the effect fading; with DNa02 silenced the bias pins at −1.000, the fly runs a fixed open-loop motor program, and deterministic physics does the rest. Since the bridge reads only DNa02, removing it necessarily removes every path from odour to motor. This test could not have failed, which makes it a consistency check rather than evidence.

M5b — the control that could fail

Cross the sensory wiring: the left antenna's odour drives the right ORN population and vice versa. Nothing downstream changes — same connectome, same DNa02 readout, same bridge constants, same calibration. The fly still walks and still responds to odour, so this control can fail while producing behaviour.

If the connectome is genuinely transducing the lateralisation, DNa02's ipsiversive response must now point the wrong way and the mirror difference must invert.

	mirror delta
intact	+14.15 mm	baseline
crossed wiring	+0.28 mm	98% reduction

Partial. Crossing the wiring destroys the steering — which rules out the worst case, that the bridge manufactures the behaviour on its own. But it does not cleanly invert it, which is what strong evidence would look like. The crossed condition's variance is large (srcR = [21.31, −16.23, 14.99], pooled SD 14.24), so at three seeds we cannot distinguish "abolished" from "inverted".

M5c — the statistics, because n=3 was not enough

The M5 pass was reported at n=3 per cell. Run properly, that is t = 1.98, df = 4, p ≈ 0.12 — not significant. An effect size of 1.61 says nothing about how many samples produced it. M5c adds seeds 3–5 to both the intact and crossed conditions (n=6) and runs a Welch t-test.

condition	mirror delta	95% CI	Welch t	p
intact	+11.02 mm	[+0.36, +21.68]	+2.321 (df 9.5)	0.044	significant
crossed wiring	−0.41 mm	[−16.03, +15.21]	−0.059 (df 8.9)	0.954	n.s.

Crossing the wiring removes 96% of the effect.

The result holds, but read the confidence interval. p = 0.044 is marginal and the lower bound of the CI is +0.36 mm — nearly zero. This is a real effect at the level of "publishable as preliminary," not a robust one. It rests on n=6, 2 s episodes, and contrast_gain = 55.

What the three tests establish together: the steering depends on DNa02 (M5) and on the sensory wiring being correct (M5b/M5c). The second is the load- bearing one, because unlike the silencing ablation it was free to fail — the fly still walked and still responded to odour with the wiring crossed, and the odour-dependence disappeared anyway. That is the strongest evidence here that the connectome's lateralisation, not the bridge, is carrying the signal.

What they do not establish: the crossed condition abolishes the effect rather than inverting it, so the sign is not confirmed to pass through the connectome — only that correct wiring is necessary. Distinguishing those would need more seeds or longer episodes.

---

M6 — making it fast enough to live in, and a world to live in

Every optimisation here is bit-exact and was verified before adoption. The gate is M1: FastBrain reproduces it exactly — 17,393 spikes, 403 active neurons, the same numbers, not merely in band — and is step-for-step identical to the reference over 500 steps at a fixed seed.

change	measured	exact?
int32 CSR indices	0.686 → 0.565 ms SpMV	96/96 bit-identical
ring-buffer delay line	0.0733 → 0.0110 ms	60/60 identical
memoised Fly.get_observation	3.306 → 2.743 ms/step	max|Δ| = 0.0 in qpos
render moved out of the loop	−158 ms/exchange	output unchanged
closed loop	0.0137× → 0.0257×	1.88×

Profiling overturned the obvious hypotheses

Rendering is 43% of the loop and physics 42%; the brain is 14%. So even an infinitely fast brain buys 1.16×. torch.roll — the intuitive suspect, copying 21 MB per timestep — is 1.2%, and the per-step spike recording is 0.3%. Both were "obviously" the bottleneck and neither is.

Three measurement errors worth recording, since each looked like a result:

• The first render benchmark timed skipped frames. Camera.render() early-outs unless enough simulated time has passed (camera.py:223), so a tight loop of render() calls mostly measures the early-out. Per frame actually produced it is ~158 ms and essentially independent of resolution (155.6 ms at 320×240 vs 158.7 at 640×480). I then attributed that to dm_control overhead, on the strength of a bare mujoco.Renderer managing 7.7–10.2 ms. That was also wrong — the toy benchmark used a 3-geom box-and-plane model. On the actual 71-geom mesh fly, a bare mujoco.Renderer costs 163 ms/frame too. The expense is the model's geometry, not dm_control and not rasterisation. Moving rendering off the critical path is still worth 43%; it just does not get cheaper offline, you only pay it for the frames you actually want.
• solver iterations=1000 (vs MuJoCo's default 100) looked alarming. solver_niter is 2 — Newton converges in two. Irrelevant.
• The first two attempts at the observation patch were silent no-ops: patching the instance cannot intercept super().get_observation(), and patching Simulation misses post_step, which calls Fly directly.

Refused, deliberately

FlyGym registers 2,268 collision pairs of which only 1,182 are unique — _init_self_contacts de-duplicates on the ordered key f"{geom1}_{geom2}", so both (A,B) and (B,A) are added. Removing the 1,086 duplicates would cut broad-phase work, but over 3,000 walking steps no duplicated pair was ever in contact, so the saving is unproven — and if legs ever do touch, removing one changes the contact force. That is a dynamics change, not an optimisation. Worth reporting upstream.

The world

LivingWorldArena: a wind-shaped Gaussian plume over N sources, depletable reserves with regrowth, and a toroidal odour field by minimum image — the fly is never teleported, so there is no discontinuity in the brain's input and no bound on lifetime. OdorArena already supports N sources × K dimensions; the subclass exists only because diffuse_func receives a scalar distance and therefore cannot express direction.

Two traps handled: mutating peak_odor_intensity is a silent no-op (get_olfaction reads a cache built in __init__), and calibrate_gain pins the plume's on-axis intensity to the old r⁻² value at a reference radius — without it, swapping the field would quietly rescale intensity_to_hz and move the result while looking like a pure environment change.

60 s of fly life

Path 1112.6 mm, closest approach to FOOD_B 14.8 mm from a start 83 mm away, sensed odour up 10.4×. Zero source visits — the feed radius is 3 mm.

The stronger evidence is within-run. Across 12,000 exchanges, the correlation between inter-antennal odour contrast and steering bias is:

lag(exch)    0      60     100     140     200     300
   r      -0.036  +0.106  +0.182  +0.187  +0.029  -0.089

Peak r = +0.187 at lag 140, p = 0.016 (rotation permutation — shuffling would have manufactured a p-value, since both series are autocorrelated).

…and the null control says that is not enough

The odour-blind run (--blind) drives the ORNs with the same total rate and zero left/right contrast, so odour information cannot reach the DNs. I predicted its correlation "must be ~0 by construction". It is not:

	peak |r|	at lag	p
intact	0.187	140	0.016
blind	0.082	60	0.274

The pre-registered bar was intact > 3× blind. 0.187 < 3 × 0.082, so this fails, and the approach to FOOD_B should be read as drift, not taxis.

The null exposed a confound I had not thought of, and my "structurally zero" claim was simply wrong: steering causes contrast. Turning changes which antenna faces the gradient, so bias → heading → contrast is a feedback path that exists whether or not odour drives steering. Any correlation between bias and contrast is therefore bidirectional by construction.

What remains suggestive but unproven: the intact fly's sensed odour rose 10.4× against the blind fly's 1.5×, and the lag profiles differ in shape (intact starts negative at lag 0 and rises to a delayed peak; blind starts positive and stays flat). Both are n=1 per condition. Settling it needs replicates, at ~39 min of wall clock each.

What this is not. The fly has no valence machinery: "aversive" is a sign imposed in numpy, and depletion is arena bookkeeping the brain cannot perceive except as concentration. This is a persistent environment, not foraging — and on the current evidence, not demonstrated odour tracking either.

Layout

brain.py              steppable wrapper around the v783 LIF model
bridge.py             THE BRIDGE -- every hand-tuned constant lives here
m1_sugar.py           M1 validation + benchmark
m2_body.py            M2 body, odour readout, turning, positive control
m3_neurons.py         M3 neuron identification -> neurons.v783.json
m4a_orn_dn.py         M4 go/no-go: does ORN drive lateralise the DNs?
m4_loop.py            M4 closed loop + mirrored odour-source test
m4_video.py           side-by-side video + trajectory figure
m5_ablation.py        M5 ablation with specificity control
neurons.v783.json     auditable root IDs, with source checksums
docs/hardware-notes.md      Pascal / CUDA / WSL2 findings
docs/bridge-parameters.md   every place I guessed
setup/                numbered, idempotent environment scripts
external/fly-brain/   upstream connectome model (git clone)

Setup

# Windows, elevated -- enables virtualization, installs WSL2. Then REBOOT.
powershell -ExecutionPolicy Bypass -File setup/windows/enable-wsl.ps1

# inside WSL
sudo bash setup/01-base.sh          # user, wsl.conf, toolchain, GL/EGL
sudo bash setup/02-cuda.sh          # CUDA 12.6 + sm_61 compile-and-execute gate
bash setup/03-conda-torch.sh        # miniforge, torch cu126, Pascal gate
bash setup/07-flygym.sh             # flygym 1.2.1, re-verifies the gate

Then, in order: m1_sugar.py, m2_body.py, m3_neurons.py, m4a_orn_dn.py, m4_loop.py, m5_ablation.py.

If Valorant stops working

Enabling WSL2 turns on the Hyper-V hypervisor, which Riot Vanguard may reject. You do not need to uninstall anything:

setup/windows/toggle-hypervisor.ps1 -Mode Off    # gaming; reboot
setup/windows/toggle-hypervisor.ps1 -Mode Auto   # WSL2;   reboot

---

Provenance

Connectome data is FlyWire v783 (Dorkenwald et al.; Schlegel et al.). The LIF model follows Shiu et al., Nature 2024. Respect the upstream licences and citation requirements of both.