FlyHead

FlyHead

One fruit fly's whole-brain connectome, driving a wheel-legged robot.

Take it apart.

  • Vendor mesh, details for display
  • Illustrative part
6direct-drive joints
94 mmwheel radius

Balance stays with the factory controller.

The brain only sets speed.

A real, LiDAR-mapped site.

The white line is a recorded trajectory.

45 msite span
550,975map points
25.5 slongest stuck
0.0 slongest stuck
Closed-loop simulation

Without a brain, blocked means stuck.

1.75 GiBonboard memory, too small for the brain
0.3 slongest life of any command

Distributed. The brain runs off-board.

Lose the link for 2 s and it stops.

Connectome simulation

A whole fly brain.

138,639 neurons. 15,091,983 connection records.

    Colour follows FlyWire cell-class annotations. Each point is one sampled neuron at its real coordinate. Hover a label to isolate it.

    493input cells
    • LC4104
    • LPLC2210
    • LC9179
    Engineering proxy

    LiDAR becomes neural input.

    From brain region to motion.

    Visual projection neurons→Descending neurons→Wheel-leg command

    Firing rates: Stage 1 simulation. Commands: engineering decoder v6, not reconstructed fly motor control. Signal and wheel animation illustrative.

    Real-robot record · person illustrative
    59escapes in 153 s of brain-driven running

    FlyHead

    Specs

    138,639neurons · FlyWire v783 connectome simulation
    15,091,983connection records
    493input cells
    LC4 104 · LPLC2 210 · LC9 179
    20readout neurons
    84 mscompute per 100 ms of model time
    real-robot shadow run, p50
    0.35 sloom to decision, median
    4 of 12 never triggered
    67 minlongest brain-driven run
    contacts not recorded
    216 mlongest distance in one run
    includes a non-neural exploration drive
    −0.30 m/sescape reverse command
    measured back at most 0.77 m
    6direct-drive joints

    TODO

    View the project on GitHub github.com/12sqawdwq/FlyHead →