A fly connectome
steering MicroDuck

95 / 100 held-out goals · zero falls

We coupled a fixed spiking model of the full released FAFB v783 fly connectome (139,255 neurons; 15,091,983 directed connections) to an already trained MicroDuck locomotion policy.

Goal-relative state is encoded as six sensory stimulation signals. Activity propagates through the fixed LIF network, and a trained 83,974-parameter readout maps 1,305 descending-neuron firing rates to forward, turn and stop commands. The pretrained ONNX policy controls the robot's joints. Only the readout is learned; the fly network does not learn the gait.

The selected controller reached 95/100 held-out goals with zero falls. Zero stimulation, zero readout features, and shuffled feature identities each produced 0/100 successes. An initial-pose perturbation test reached 98/100; this single result does not imply that perturbations improve performance. Success requires entering a 15 cm goal radius and remaining within 20 cm for one second, within a 30-second episode.

The video synchronizes the robot, the detailed articulated TuragaLab flybody 3D asset, and actual filtered firing rates mapped onto all 139,255 anatomical FlyWire neuron reference coordinates. The brain view uses a rotating 3D point cloud inspired by fly-brain-vis; it is not a reconstruction of neuron skeletons. The fly mesh replays robot heading and an illustrative joint gait driven by measured motion; it is not an independently controlled physical fly simulation. Robot observations are checked against the original recording at every replay step.

These controls establish dependence on neural signals for this trained controller, not an advantage over random wiring. Sensory assignments, transmitter signs (including an excitatory default for unknown labels), and discretized dynamics are modeling assumptions. This is simulated robot control, not evidence of biological equivalence or physical-robot performance.

Inspired by Shiu et al.'s fly-brain model, using FlyWire data and Pollen Robotics' MicroDuck simulator.