fruitfly.arnao.ai — research dossier (2026-09-25)

Two background research passes (Sonnet-class agents, every URL verified live at fetch time). Part A = the science and data landscape; Part B = browser feasibility.

---

Part A — The "AI fruit fly" landscape

TL;DR. Between Oct 2024 and Sep 2026 the fly went from "mapped" to "runnable": a whole-brain female connectome (FlyWire, 139k neurons), a whole-CNS male connectome (MaleCNS v1.0, 166.7k neurons, Cell, 2026-09-03), a female brain+cord connectome (BANC, ~188k neurons, Nature, 2026-06-08), two physics bodies (DeepMind/Janelia flybody; EPFL NeuroMechFly v2), and a whole-brain spiking model (Shiu et al.) that Eon Systems wired into a body in March 2026. In September 2026 this exploded into a viral hobbyist wave (Beat Saber clip, 22M views) with 129 catalogued community projects, ~35 of them browser-playable. Almost all browser demos are LIF-brain "poke a neuron" toys; none combines an accurate 3D body, a physically-simulated MuJoCo fly, the real neuron meshes, and real behavior data in one casual, no-login experience.

1. flybody — DeepMind + Janelia whole-body MuJoCo fly

• Paper: Vaxenburg, Siwanowicz, Merel, Robie, Morrow, Novati, Stefanidi, Both, Card, Reiser, Botvinick, Branson, Tassa, Turaga. Whole-body physics simulation of fruit fly locomotion, Nature 643:1312–1320, published 2025-04-23. https://www.nature.com/articles/s41586-025-09029-4 (open PMC copy: https://pmc.ncbi.nlm.nih.gov/articles/pmid/40267984). Preprint 2024-03-11: https://www.biorxiv.org/content/10.1101/2024.03.11.584515v2
• Code/model: https://github.com/TuragaLab/flybody — Apache-2.0. Contains the MJCF body model, RL task envs (walking imitation, flight, vision-guided flight), Ray/DMPO training scripts, tutorial notebooks. No trained policies or datasets in the repo itself.
• Mirror in MuJoCo Menagerie (canonical MJCF, Apache-2.0, "requires MuJoCo 2.2.2+"): https://github.com/google-deepmind/mujoco_menagerie/tree/main/flybody
• Public data (Janelia Figshare, posted 2025-05-07, license reported by Figshare API as GPL-3.0+, ~4.05 GB total, anonymous download): https://janelia.figshare.com/articles/dataset/MuJoCo_fruit_fly_body_model_datasets_supporting_Whole-body_simulation_of_realistic_fruit_fly_locomotion_with_deep_reinforcement_learning_/25309105
• datasets_walking-imitation.zip — 3.02 GB (~16,000 walking snippets, ~80 min of real fly walking as full-body joint angles/positions/velocities)
• dataset_fly-confocal-stack.zip — 1.24 GB (anatomy source)
• datasets_flight-imitation.zip — 12.9 MB (flight reference trajectories)
• trained-fly-policies.zip — 6.5 MB; flight-controller-reuse-checkpoints.zip — 32 MB
• videos_fly-anatomy.zip — 37 MB; dataset_fly-grooming-poses.zip — 0.3 MB

(File list from the public API: https://api.figshare.com/v2/articles/25309105)

• Body stats as consumed downstream: 67 body parts, 111 muscles/actuators (per webgpu-fly's use of the model, https://webgpu-fly.pages.dev/). Also wrapped by FlyGym: https://neuromechfly.org/api_reference/flygym/compose/fly/flybody/
• 2026 follow-ons: FlyGM (Tsinghua/Georgia Tech, arXiv 2026-02-20) uses the whole-brain connectome graph as an RL controller for the flybody: https://arxiv.org/abs/2602.17997 (NeurIPS 2025 workshop version: https://neurips.cc/virtual/2025/131402). GPU: MuJoCo Warp https://github.com/google-deepmind/mujoco_warp

2. FlyWire whole-brain connectome (female, FAFB)

• Paper: Dorkenwald et al., Neuronal wiring diagram of an adult brain, Nature, 2024-10-02 (nine-paper collection): https://www.nature.com/collections/hgcfafejia ; overview https://www.nature.com/articles/s41583-024-00879-x
• Scale: 139,255 proofread neurons, ~50M synapses (54.5M per https://www.sciencenews.org/article/fruit-fly-brain-connections-traced), current release v783 (Oct 2023 snapshot).
• Portal: https://flywire.ai/ ; apps page https://flywire.ai/apps
• Codex (Connectome Data Explorer): https://codex.flywire.ai/ — FAQ https://codex.flywire.ai/faq confirms: interactive apps require Google sign-in (rate-limiting); static pages don't. Downloads are gzipped CSV: neurons.csv.gz, connections.csv.gz, classification.csv.gz, consolidated_cell_types.csv.gz; SWC skeletons per cell where available; meshes are NOT served by Codex — use MeshParty/CloudVolume. Hosted datasets: FAFB v783, BANC v626/v888, MANC v1.2.1, MAOL v1.1, MCNS v0.9/v1.0. Codex source: https://github.com/murthylab/codex
• Neuroglancer: public instance https://ngl.flywire.ai/ . Public flat v783 segmentation + meshes in gs://flywire_v141_m783 readable without credentials (per https://github.com/Flux-Frontiers/connectome_kg/pull/12); URL helpers: https://fafbseg-py.readthedocs.io/en/latest/source/tutorials/neuroglancer.html
• CAVE: https://github.com/CAVEconnectome/CAVEclient ; access guide from Seung lab: https://github.com/seung-lab/FlyConnectome (Codex, CAVE, CloudVolume, braincircuits.io, CATMAID)
• Community annotations (CSV/parquet): https://github.com/flyconnectome/flywire_annotations
• License: the Nature data release and the Zenodo skeleton release are CC-BY 4.0 (https://zenodo.org/records/10877326); the FlyWire ToS says edits/annotations are released CC-BY-NC 4.0 (non-commercial): https://flywire.ai/tos. MANC/MaleCNS are CC-BY. Cite per https://flywire.ai/guidelines
• Compact packaged copies: Shiu's repo ships v630 and v783 connectivity as parquet (~100 MB): https://github.com/philshiu/Drosophila_brain_model ; fly-explorer's v783 export is 32 MB: https://github.com/brandoncho369/fly-explorer

3. Connectome-constrained models

4. Brain-in-body: Eon, NeuroMechFly, MANC, BANC, MaleCNS, 2026 news

• Eon Systems (SF), 2026-03-10: "How the Eon Team Produced a Virtual Embodied Fly" — Harris, Sinha, Toth, Pomares, Shiu; co-founder Alex Wissner-Gross. Shiu LIF brain (~140k neurons/50M synapses) + flyvis eye + NeuroMechFly v2 body (87 joints) in MuJoCo; behaviors: grooming, feeding, foraging, looming-escape circuit modeled. No code/data released; video only. https://eon.systems/updates/embodied-brain-emulation ; coverage https://the-decoder.com/startup-claims-first-full-brain-emulation-of-a-fruit-fly-in-a-simulated-body/ , https://www.theregister.com/offbeat/2026/03/16/digital-fruit-fly-brain-model-walks-and-cleans-its-feelers/5224139 ; next target: mouse within two years.
• NeuroMechFly v2 / FlyGym (EPFL Ramdya lab): Wang-Chen et al., Nature Methods 21:2353 (2024) https://www.nature.com/articles/s41592-024-02497-y ; code https://github.com/NeLy-EPFL/flygym (Apache-2.0; FlyGym 2.x, March 2026 rewrite, ~300x GPU speedup via MJWarp); docs https://neuromechfly.org/ . Review "The embodied brain" (Jan 2026): https://arxiv.org/pdf/2601.08056
• MANC (male adult nerve cord): Takemura et al. 2024; ~23k neurons. neuPrint https://neuprint.janelia.org ; project page https://www.janelia.org/project-team/flyem/manc-connectome ; CC-BY, flat files on a public Google bucket; R: https://github.com/natverse/malevnc
• MaleCNS (whole male CNS): v1.0 2026-06-08; 166,700 neurons, 11,691 cell types; Cell paper Berg et al., 2026-09-03: https://www.cell.com/cell/fulltext/S0092-8674(26)00942-6 (preprint https://www.biorxiv.org/content/10.1101/2025.10.09.680999v1). Site https://male-cns.janelia.org/ ; downloads https://male-cns.janelia.org/download/ — CC-BY, no login for bulk: Feather tables (connection weights 1.1 GB, synapse points 12.7 GB, partners 6.8 GB), SWC + neuroglancer-precomputed skeletons. Janelia news: https://www.janelia.org/news/researchers-reveal-connectome-of-the-male-fruit-fly-central-nervous-system
• BANC (female brain-and-nerve-cord): Bates, Phelps, Kim, Yang … Murthy, Drugowitsch, Wilson, Lee — Distributed control circuits across a brain-and-cord connectome, Nature 2026-06-08 (open access): https://www.nature.com/articles/s41586-026-10735-w (PMC https://pmc.ncbi.nlm.nih.gov/articles/PMC13518251/). ~188k neurons, 199M predicted synapses. Access https://flywire.ai/banc_access , Codex https://codex.flywire.ai/?dataset=banc , notebooks https://github.com/htem/BANC-project , R https://github.com/natverse/bancr. Press: https://www.sciencedaily.com/releases/2026/06/260610003047.htm
• Sept 2026 viral wave (timeline at https://knowyourmeme.com/memes/fly-brain-simulations-fruit-fly-brain-mapped): Minecraft (Ro0oney, Sep 7, 241k views), Beat Saber (@_lyraaaa_, Sep 8–9, ~22M views), Bitcoin-trading fly (Sep 10, 3.4M). Explainer: https://sg.news.yahoo.com/whats-fly-brain-simulations-viral-144806619.html
• State of Brain Emulation Report 2025 (Zanichelli, Schons, Freeman, Shiu, Arkhipov, Oct 2025): https://arxiv.org/abs/2510.15745

5. Public real-behavior datasets

6. Existing interactive visualizers

Gap analysis — what does not yet exist as a public, browser-based, visual experience

1. A "one fly, three views" scene: real FlyWire/MaleCNS neuron meshes (not point clouds) inside the flybody MJCF body, with the body walking on the Figshare real-walking trajectories. Every current demo picks one: meshes (Codex/VFB), LIF dots (fly-explorer/FlyBrain.app), or physics (webgpu-fly/Lulzx). None loads without login and shows morphology and moves.
2. Real-data playback, not RL theater. No site lets a visitor scrub the 80-minute Janelia walking corpus or the Dickinson flight kinematics on a 3D fly and compare to the trained policy.
3. Honest provenance overlay. No visualizer colors each behavior by which paper/dataset supports it vs "engineered controller" (Fly Arena, Eon, Neural Canvas all admit illustrative motion).
4. Mobile-first / under 30 MB entry. webgpu-fly is ~300 MB and WebGPU-only; Lulzx is 23 MB but desktop-recommended. A progressive tier with WebGL2 fallback does not exist.
5. Female vs male side-by-side in one interactive brain. Nobody renders the 138 dimorphic / 289 male-specific / 71 female-specific types (Berg et al.) as a togglable 3D diff.
6. Brain→cord→muscle path tracing you can drive. No demo lets a casual user click a descending neuron and watch the cascade reach specific motor neurons and then a specific body actuator, with the anatomical path drawn.
7. Benchmark-as-game. flybench has 36 tasks and a leaderboard but no visual front end.
8. Embeddable, shareable states. Only fly-explorer encodes experiments in URLs. An embeddable component with licensing baked in is missing.

Licensing note for a new site: FlyWire data CC-BY 4.0 with annotations CC-BY-NC 4.0 (https://flywire.ai/tos); MaleCNS and MANC are CC-BY (https://male-cns.janelia.org/download/); flybody code/MJCF is Apache-2.0 (https://github.com/TuragaLab/flybody) but its Figshare datasets are tagged GPL-3.0+ (https://api.figshare.com/v2/articles/25309105); flyvis and Shiu's model are MIT; FlyGym Apache-2.0.

Unverified items: BrainCircuits.io not fetched live; the Seung-lab Aug-2026 preprint's code link could not be read (bioRxiv rate-limited); Figshare per-file contents inferred from names and the Nature methods text.

---

Part B — Browser feasibility

Bottom line: every piece already runs in a browser somewhere. Google DeepMind ships official MuJoCo WASM bindings, flybody's assets and trained policies are downloadable and small, FlyWire data is CC-BY 4.0 with public GCS buckets, and at least four independent projects already run the full 139k-neuron LIF brain (some with the MuJoCo fly body) client-side. The open work is integration, asset packaging, and mobile fallbacks, not research risk.

1. flybody in the browser (MuJoCo WASM)

• Official bindings exist. DeepMind publishes @mujoco/mujoco on npm; versions follow MuJoCo releases. Single-threaded build works everywhere; /mt build uses Web Workers + SharedArrayBuffer and requires COOP/COEP headers. Status "WIP". https://github.com/google-deepmind/mujoco/blob/main/wasm/README.md
• Current version 3.14.0, Apache-2.0, unpacked 22.1 MB. https://registry.npmjs.org/@mujoco%2Fmujoco/latest
• zalo/mujoco_wasm is now a demo suite over the official bindings; shows MJCF + three.js rendering. https://github.com/zalo/mujoco_wasm and https://zalo.github.io/mujoco_wasm/
• MJCF with meshes: bindings load MJCF via MjModel.from_xml_string(); rendering is yours (three.js). idiap fork adds explicit mesh handling. https://github.com/idiap/mujoco_wasm3
• flybody assets: 142 .obj meshes totalling ~172 MB raw, plus fruitfly.xml (66 KB). https://api.github.com/repos/TuragaLab/flybody/contents/flybody/fruitfly/assets. Paper: simplified geometry ~20,000 faces, 67 bodies, 102 DoF, physics timestep 0.05 ms (flight) / 0.2 ms (walking). https://pmc.ncbi.nlm.nih.gov/articles/PMC12310536/ — decimated + meshopt-compressed glTF should land under ~10 MB.
• Proof it works: webgpu-fly runs the flybody model in-tab (~300 MB first visit, 0.25 kHz biological time on an M2 Pro). https://webgpu-fly.pages.dev/ ; Lulzx/fly-brain also runs flybody in-browser (viewer ~30 MB). https://github.com/Lulzx/fly-brain

2. Trained flybody policies

• Public checkpoints via download_data.py: https://raw.githubusercontent.com/TuragaLab/flybody/main/flybody/download_data.py
• Sizes: trained-fly-policies.zip 6.5 MB; flight-controller-reuse-checkpoints.zip 32 MB; walking-imitation 3.0 GB; flight-imitation 12.9 MB. Figshare lists the dataset license as GPL-3.0+ (repo Apache-2.0; paper CC-BY). https://api.figshare.com/v2/articles/25309105 — clarify with Turaga lab before redistribution.
• Framework: TensorFlow + Sonnet + Acme (DMPO). Policy = LayerNormMLP(256,256,256) → MultivariateNormalDiagHead. No LSTM. https://raw.githubusercontent.com/TuragaLab/flybody/main/flybody/agents/network_factory.py. Vision-flight adds a small CNN. https://raw.githubusercontent.com/TuragaLab/flybody/main/flybody/agents/network_factory_vis.py
• I/O dims: walking obs 286 → action 59; flight obs 62 → action 12. https://pmc.ncbi.nlm.nih.gov/articles/PMC12310536/
• Loading the pretrained walker needs "TF 2.8 + dm-reverb, Linux x86 only." https://github.com/Chere3/flyphone/issues/2
• Browser path: a 3×256 MLP is ~200k params (<1 MB fp32). Dump weights to Float32Array and hand-write the forward pass (~60 lines of JS); keep ONNX (https://github.com/onnx/tensorflow-onnx, onnxruntime-web 1.30.0 https://registry.npmjs.org/onnxruntime-web/latest) for the vision policy only.

3. FlyWire connectome in the browser

• License: FAFB v783 CC-BY 4.0 (Dorkenwald et al. 2024). https://www.nature.com/collections/hgcfafejia ; skeleton release on Zenodo CC-BY 4.0. https://zenodo.org/records/10877326
• Skeletons: sk_lod1_783_healed_ds2.parquet = 5.4 GB for all neurons. https://zenodo.org/records/10877326. A full-res mesh example has ~310k vertices / 623k faces, so LOD/decimation is mandatory. https://fafbseg-py.readthedocs.io/en/latest/source/tutorials/flywire_neurons.html
• Meshes: Codex does not serve meshes. https://codex.flywire.ai/faq. Public precomputed buckets: flat v783 segmentation at gs://flywire_v141_m783, neuropils at gs://flywire_neuropil_meshes. https://github.com/Flux-Frontiers/connectome_kg/pull/12 ; https://github.com/sjcabs/fly_connectome_data_tutorial. Neuroglancer's multi-resolution precomputed mesh format is Draco-encoded, LOD'd, CDN-friendly. https://github.com/google/neuroglancer/blob/master/src/datasource/precomputed/meshes.md
• Synapses: full per-synapse table 9.5 GB (~130M rows); pair×neuropil table 852 MB. https://zenodo.org/records/10676866
• API/CORS: Codex apps require Google sign-in, no bulk API. https://codex.flywire.ai/faq. CAVE tokens needed. https://caveclient.readthedocs.io/en/latest/guide/authentication.html. Conclusion: pre-bake a static subset; never hit CAVE/Codex from the browser.
• Existing three.js viewers: fly-brain-bench (three.js r180, whole graph 30.7 MB gzipped, filtered ≥5-synapse graph 6.8 MB) https://github.com/khpnhfpkpz-max/fly-brain-bench and https://flybrain.raphaelrocha.com/ ; snedea/flybrain https://github.com/snedea/flybrain ; Lulzx https://lulzx.com/fly-brain/ ; list: https://github.com/cobanov/awesome-fly

4. Shiu et al. 2024 LIF model

• Repo: MIT, Brian2, Connectivity_783.parquet = 100.8 MB. https://github.com/philshiu/Drosophila_brain_model
• Model: 127,400 neurons, 50M synapses; V_rest = V_reset = −52 mV, threshold −45 mV, refractory 2.2 ms, synaptic τ 5 ms, delay 1.8 ms, W_syn 0.275 mV; Brian2 ~5 min per 1,000 ms trial per CPU thread. https://pmc.ncbi.nlm.nih.gov/articles/PMC11446845/
• Browser feasibility (proven): fly-brain-bench runs 138,639 neurons at 60 fps in plain JS via an active-set scheduler (~126k spikes/s). https://github.com/khpnhfpkpz-max/fly-brain-bench. webgpu-fly runs the same model as fused WebGPU compute kernels. https://webgpu-fly.pages.dev/. fly-connectome-webgpu runs MaleCNS (165k neurons, 25.6M edges) with TensorFlow.js WebGPU. https://github.com/itsyuimorii/fly-connectome-webgpu
• Size budget: thresholded edges ≈ 7–30 MB compressed; full 50M-synapse set ~100 MB; ship the thresholded graph by default.

5. Vercel hosting constraints

• CLI source upload cap 100 MB (Hobby) / 1 GB (Pro). https://vercel.com/docs/limits
• "Do not package large videos, archives, or changing data exports into each deployment." https://vercel.com/docs/deployment-storage/optimize
• Vercel Blob: CDN + browser cache up to 1 month; blobs > 512 MB are never cached. https://vercel.com/docs/vercel-blob/public-storage and https://vercel.com/docs/vercel-blob/usage-and-pricing
• Hobby: 1 GB storage, 10 GB Blob transfer, non-commercial only. Pro is usage-based (~$0.05/GB transfer). https://vercel.com/docs/limits/fair-use-guidelines
• Streaming: chunk to ≤ 50 MB files; meshopt (EXT_meshopt_compression) over Draco for neuron meshes. https://meshoptimizer.org/gltf/ ; https://github.com/mrdoob/three.js/pull/20508

6. Libraries (versions verified 2026-09-25)

WebGPU 2026: Chrome 113+ desktop / 121+ Android; Firefox 141+ Windows, 145–147+ macOS; Safari 26+ on macOS and iOS. https://github.com/gpuweb/gpuweb/wiki/Implementation-Status ; caniuse global ~87%. https://caniuse.com/webgpu. Firefox Linux/Android still needs a WebGL2/Worker fallback.

Recommended architecture

1. Static site on Vercel (Pro for the 1 GB upload headroom and commercial-safe terms); app bundle under 10 MB including the single-threaded MuJoCo WASM.
2. Assets on a public Vercel Blob store, every object under 512 MB, content-hashed immutable names: decimated flybody glTF (meshopt) ~5–10 MB; fruitfly.xml; policy weights as .bin (<2 MB); connectome subset — 1,000 curated neurons as LOD'd meshopt glTF (~50–150 MB total, lazy per neuron) plus a whole-brain soma point cloud (~1.7 MB); thresholded CSR connectome (~7–30 MB); one or two walking reference trajectories from the 3 GB imitation dataset.
3. Client compute: MuJoCo in a Web Worker, policy MLP hand-ported to JS, LIF brain in a second Worker with an active-set scheduler; optional WebGPU compute path. Render via three.js WebGL2.
4. Fallbacks: WebGL2 rendering everywhere; JS-Worker LIF where WebGPU is absent; single-threaded MuJoCo unless COOP/COEP is set; reduced neuron count on phones.
5. No live FlyWire/CAVE calls; all data pre-baked by a Python pipeline (fafbseg/navis) run on the Mini.

Risks and mitigations

• Flight physics too slow in WASM (0.05 ms timestep). Default to walking; flight sub-real-time with a speed indicator; try MT build behind COOP/COEP.
• Checkpoint license ambiguity (Figshare GPL-3.0+, repo Apache-2.0, paper CC-BY). Email Turaga lab; worst case retrain a walking policy with MJX.
• TFP distribution head fails tf2onnx. Export weights only; MLP forward pass in JS.
• COEP blocks cross-origin Blob fetches. Verify with curl; use credentialless or same-origin /data/* rewrite; default to the ST build.
• Blob egress cost from hot-linking / bots. Blob firewall rate-limit rules; keep default payload under 30 MB.
• Memory on mobile. LOD, per-neuron lazy load, device-tier gating.
• FlyWire bucket access assumptions (anonymous GCS reads reported in a PR, not official docs). Verify before building the pipeline; fall back to the Zenodo skeleton parquet.
• WebGPU gaps. JS/WebGL2 path is the baseline, WebGPU is progressive enhancement.
• Scientific over-claim. webgpu-fly's own caveat: "teaching and intuition-building tool, not a validated scientific platform." Disclosure copy on the site.