FlyCNS Embodied Control and MaleCNS Spectral Mixing in Drosophila

Once nanoscale electron microscopy maps every neuron and chemical synapse across an entire brain and nerve cord, what can that static wiring matrix actually do on its own, and where does pure topology hit a wall without physiology? Two preprints posted in the final week of September 2026 put the adult Drosophila melanogaster central nervous system connectome through opposite, complementary tests.

In “FlyCNS: Connectome-Grounded Dynamic Sparse Communication for Embodied Control” (arXiv:2609.28816), Jinchang Zhang, Zijian Lin, and Guoyu Lu at the University of Georgia extract a macro-scale directional communication prior from the Drosophila brain-and-nerve-cord connectome and embed it inside NeuroRoute, a Lagrangian sparse communication controller running on a physical 12-joint Unitree Go1 quadruped. In “A Pre-Registered Spectral and Certified Mixing Analysis of the Adult Male Drosophila Central Nervous System Connectome” (arXiv:2609.33054), Eran Kopel at Tel Aviv University runs five locked, pre-registered mathematical hypotheses on the complete Janelia MaleCNS v1.0 graph of 166,700 neurons and 124,177,617 synapses, testing whether raw synapse-weighted random walks actually integrate information across the central nervous system.

Taken together, the two studies sharpen the boundary between anatomical structure and functional dynamics that we track on the Brain Emulation section, in our earlier comparison of MaleCNS v1.0 and FlyWire, and in the review of Eon Systems’ fruit fly brain emulation. Macro-scale connectome pathway ratios supply a highly effective structural prior for embodied sensorimotor bandwidth allocation, while raw synapse-count transition matrices without physiological gain control and post-inhibitory rebound get trapped at peripheral motor boundaries.

Part 1 FlyCNS and NeuroRoute on a 12-Joint Quadruped

Embodied robots and biological nervous systems share a strict engineering constraint. Peripheral limbs generate continuous high-dimensional proprioceptive streams (ascending traffic), while central coordination modules issue motor commands (descending traffic) across a neck connective or communication bus of limited bandwidth. Standard multi-agent or modular robot controllers either broadcast full state vectors at every step (FULL), drop packets on a fixed schedule (PERIODIC), or learn unstructured gates from scratch.

Zhang, Lin, and Lu ask whether the evolutionary wiring of the fruit fly central nervous system encodes a reusable bandwidth-allocation law. From the adult Drosophila brain-and-nerve-cord connectome, they coarse-grain the inter-regional pathway matrices connecting sensory/ascending clusters and motor/descending clusters into an ascending effective matrix $W^\uparrow \in \mathbb{R}^{3 \times 16}$ and a descending effective matrix $W^\downarrow \in \mathbb{R}^{16 \times 3}$. Using the Shannon entropy of the singular-value spectrum of each pathway block, they compute the effective rank $\mathrm{erank}(W)$ of the biological pathways and normalize them into a directional structural prior $(\tilde{p}^\uparrow, \tilde{p}^\downarrow)$.

Instead of hardcoding fixed gates, FlyCNS feeds $(\tilde{p}^\uparrow, \tilde{p}^\downarrow)$ into NeuroRoute, a primal-dual Lagrangian actor-critic architecture controlling a 4-limb, 12-joint Unitree Go1 quadruped trained over 206.4 million environment interactions (50,400 updates). Full inter-module communication on the quadruped requires $C_{\mathrm{full}} = 8,224\text{ bits/step}$. Under a strict 25% bandwidth target ($B = 0.25$), NeuroRoute dynamically gates ascending limb-to-coordinator messages and descending coordinator-to-limb messages at every control step while penalizing deviation from the connectome’s directional effective-rank ratio.

The result is striking. Using less than 22.2% of the communication bus (1,755 to 1,823 bits per step), FlyCNS holds a tracking score of 0.882 across both Official and Scripted locomotion command profiles, staying within 6.1% of unconstrained full-bandwidth control while PERIODIC (0.5471) and unguided ASYM (0.4359) collapse. Furthermore, inspecting the learned gates across the gait cycle shows that NeuroRoute spontaneously concentrates ascending communication during stance-to-swing transitions when ground-reaction surprise is highest, demonstrating how a static connectomic effective-rank prior can scaffold real-time event-driven routing.

Part 2 Eran Kopel’s Pre-Registered Audit of MaleCNS v1.0

Where FlyCNS extracts a coarse-grained macro prior across pathway blocks, Eran Kopel’s study (arXiv:2609.33054) tests the fine-grained graph of MaleCNS v1.0 (166,700 neurons, 124,177,617 synapses across 25,582,938 directed edges). In the spirit of the adversarial collaborations and measurement audits we cover in the 2026 consciousness test survey, Kopel locked five mathematical hypotheses (H1 through H5) in a pre-registered protocol before running the spectral and Markov-chain mixing analysis on the giant strongly connected component (SCC).

At raw threshold $w \ge 1$, the giant SCC $G$ contains 165,314 neurons (99.17% of the CNS); at the noise-filtered threshold $w \ge 5$, the giant SCC contains 157,821 neurons. Kopel models signal propagation as a row-normalized random walk $P(F){ij} = \frac{w{ij}}{s_{\mathrm{in}}(i) + F}$, where $F \ge 0$ is an input-degree regularization floor that prevents peripheral neurons with tiny input synapse counts from dominating the spectrum.

What Passed (H2 and H5) Filtering Out Photoreceptor Artifacts to Reveal the Ellipsoid Body Ring

In unregularized ($F = 0$) spectral analyses of fly connectomes, the leading slow eigenvectors are routinely dominated by retinal photoreceptor terminals (R7 and R8), which have very small input degrees inside the optic lobe and therefore act as artificial sinks/sources under naive degree normalization.

Kopel’s Hypothesis H2 (that photoreceptor R7/R8 dominance is a normalization artifact eliminated by setting the regularization floor to the median input strength $F = 342$) and Hypothesis H5 (that degree regularization uncovers genuine central-complex structural modes) both passed. Once $F = 342$ suppresses peripheral low-degree inflation, the artificial photoreceptor modes vanish and a tightly localized structural mode emerges in the central complex’s ellipsoid body, carried by 20 contralateral GABAergic ring neurons of types ER3d_b (12 neurons) and ER3p_a (8 neurons) with an effective participation ratio of 96.6. In the living fly, these ellipsoid-body ring neurons form the core recurrent compass and sleep-homeostatic ring attractor of the brain.

What Failed (H1, H3, and H4) The 11-Neuron Ventral Nerve Cord Boundary Trap

By contrast, Kopel’s three global mixing hypotheses (H1, H3, and H4), which predicted that the strongly connected CNS graph would exhibit a wide spectral gap and rapid global Markov mixing across 128 synaptic steps, all failed decisively.

Why does a 157,821-neuron strongly connected central nervous system fail to mix at 128 steps under a synapse-weighted random walk? Kopel traced the bottleneck to an 11-neuron near-closed boundary set in the ventral nerve cord (VNC), anchored by seven mesothoracic efferent and neuromodulatory neurons (EN00B001, EN00B008, EN00B011, EN00B015, and mesVUM-MJ). Anatomically, these motor-output neurons receive heavy descending and local premotor drive inside the VNC, but send almost all of their output out to the wing and leg muscles, retaining only four weak internal return edges back into the rest of the SCC.

In a pure graph random walk where probability mass flows in proportion to static synapse counts, any trajectory that enters that 11-neuron mesothoracic efferent pocket is almost completely trapped within the SCC, collapsing the global spectral gap to 0.0035. In the living animal, of course, signals do not get trapped in motor efferents, because motor spikes exit the nervous system to move the body in the physical world, changing proprioceptive feedback through sensory afferents, while internal central-pattern oscillators and post-inhibitory rebound reset the circuit.

Why Both Results Are Decisive for Consciousness and Emulation

Reading FlyCNS (arXiv:2609.28816) alongside Kopel’s MaleCNS audit (arXiv:2609.33054) provides the clearest empirical proof to date of two rules we emphasize across our connectomics coverage, including our analysis of larval zebrafish spinal circuit coordination.

1. A connectome is an open sensorimotor loop, not a closed Markov chain. When you treat an anatomical graph as a closed matrix without a body (Kopel H1/H3/H4), efferent motor boundaries look like pathological probability sinks ($1 -

   \lambda_2

   = 0.0035$). When you close the loop through an embodied physical plant (FlyCNS on the Unitree Go1), the ascending-to-descending pathway ratio becomes an asset that slashes internal communication bandwidth by 78%.

2. Topological integration metrics ($\Phi$, spectral gap, Dobrushin mixing) cannot be computed naively on raw synapse counts. Without degree-floor regularization ($F = 342$ to expose the ER3d_b / ER3p_a ellipsoid-body ring) and explicit physiological dynamics (excitatory vs GABAergic sign, membrane time constants, and embodied sensory return), graph-level complexity numbers measure peripheral degree artifacts rather than central integrative capacity.

You can inspect these exact structural pathways in our interactive Connectome Atlas.

Open the Connectome Atlas Explore the biological connectome pathways discussed here in the Connectome Atlas, and compare static synaptic topology with active spiking dynamics.

Comparison to The Consciousness AI

Both papers validate design choices at the core of The Consciousness AI, documented on our Core Architecture page and interactive Substrate Console.

First, Kopel’s discovery that regularizing peripheral degree artifacts uncovers the 20-neuron GABAergic ellipsoid-body ring (ER3d_b / ER3p_a) while raw random walks fail to mix without oscillatory and embodied reset mirrors what we found in our own Layer 1 Leaky Integrate-and-Fire (LIF) and Layer 2 oscillatory binding audits. In our codebase’s rhythm binding gate (tcai-research #59 and #61), static connectivity without regular oscillatory timing fails to bind features (67/120 with rhythm destroyed vs 115/120 across 3/3 seeds with phase-locked rhythm intact). Second, Kopel’s 11-neuron efferent boundary trap is the exact biological counterpart to closing the sensorimotor loop in our tectal motor map gate (tcai-research #58, 95 and 102 of 120 target approaches), demonstrating that descending motor pathways only integrate back into the central workspace when they move an embodied agent through an environment and return as new ascending sensory input. Finally, FlyCNS’s Lagrangian effective-rank prior ($\tilde{p}^\uparrow, \tilde{p}^\downarrow$) offers a clean mathematical constraint we can test directly on the ascending-to-descending bandwidth budget of our own thalamic_relay gate.

Sources

• Jinchang Zhang, Zijian Lin, and Guoyu Lu. FlyCNS: Connectome-Grounded Dynamic Sparse Communication for Embodied Control. arXiv, posted September 24, 2026 (arXiv:2609.28816v1).
• Eran Kopel. A Pre-Registered Spectral and Certified Mixing Analysis of the Adult Male Drosophila Central Nervous System Connectome. arXiv, posted September 29, 2026 (arXiv:2609.33054v1).
• Stuart Berg, Isabella R. Beckett, Marta Costa, Philipp Schlegel, Gregory Jefferis, and colleagues. MaleCNS v1.0 dataset and sexual dimorphism in the complete connectome of the Drosophila male central nervous system. HHMI Janelia / Cell, 2026.
• The Consciousness AI Project Repository. Tectal motor map (tcai-research #58) and rhythm binding (#59, #61) verification records.

Researchers covered here

• GL Guoyu LuSchool of Electrical and Computer Engineering, University of GeorgiaFlyCNS connectome-grounded directional communication prior and NeuroRoute dynamic sparse routing for embodied quadruped locomotion 1 article
• JZ Jinchang ZhangSchool of Electrical and Computer Engineering, University of GeorgiaFlyCNS connectome-grounded embodied control and NeuroRoute Lagrangian sparse communication for quadruped locomotion 1 article
• EK Eran KopelFaculty of Medical and Health Sciences, Tel Aviv UniversityPre-registered spectral and certified Dobrushin mixing analysis of the adult male Drosophila central nervous system connectome (MaleCNS v1.0) 1 article