I connected the FAFB v783 connectome — as the LIF model of Shiu et al. 2024 — to a small bipedal robot, Pollen Robotics’ open-source microduck, in closed loop in its MuJoCo twin. Its depth sensor drives LPLC2/LC4, its camera drives T4/T5, and descending neurons move it. Nothing is trained, and the boundary between the connectome and my glue code is spelled out.
A few things from the wiring that may interest people here:
- The column map’s orientation, read off the wiring. The visual columns give hex coordinates but not which way is front or up. LPLC2’s four arms take outward motion, so its T4a/T5a inputs sit behind it and T4c/T5c above it. That fixes the axes in both eyes, and HSN vs HSS confirms up.
- A narrow camera is not enough for DNa02. Driving only the ~1,400 T4/T5 that look into a 45° × 73° view, DNp15 (DNHS1) still flips with rotation; DNa02 does not. Over the whole eye both do.
- LC10a → AOTU019/025 → DNa02 turns toward a small target, but only below the eye’s equator, and more weakly to the right: the two eyes’ LC10a receptive fields (from their column-mapped inputs) are not mirror images.
- A negative result: in this LIF model, any input strong enough to reach the mushroom body fires the same ~3,400 of 5,177 Kenyon cells (odours, taste, temperature; overlap 0.98–1.00), a runaway that starts in the antennal lobe.
Controls: with the connectome rewired (degree-preserving), the startle, the head reflex and the following all disappear.
Code and paper: GitHub - first-it-consulting/duckophila: A fruit fly's connectome driving a microduck: the FlyWire whole-brain model in closed loop with Pollen Robotics' / Hugging Face's open-source microduck · GitHub
Video: Duckophila - a Hugging Face Space by HW01
Feedback very welcome, especially on the column-to-angle mapping, and on whether the LC10a asymmetry and the mushroom-body runaway are known.
Credit: FlyWire (Dorkenwald, Schlegel, Matsliah et al. 2024), Shiu et al. 2024.