← The index · Brain models & bodies · by ben-gitdev

FLYNN

FlyWire FAFB v783 as a ~139k-unit sparse RNN trained with DAgger to drive a wheeled MuJoCo robot, vs small-world and CNN baselines. Wang & Chen, arXiv:2607.00025.

FlyWireMuJoCoRL / imitation
Fidelity · reviewed
L1 Dynamics

Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.

Dispute this rating →

The evidence

No evidence, no level.
Rating note

Every unit is one FlyWire FAFB v783 neuron and every recurrent weight one synapse-count-weighted connection, run as a leaky rate unit h_new = (1-alpha)*h + alpha*phi(W@h + b) with a per-cell-type leak rate, plus an explicit virtual-retina front end. That documented unit dynamics earns L1. Not L2 even though the connectome weights themselves are trainable: the change is driven by torch.optim.Adam minimising a DAgger imitation loss against a VFH*+PID planner, not by any modelled neuromodulator or fly learning rule, and it happens in an offline training script that saves a checkpoint - at rollout time the weights are fixed. Gradient descent on a connectome-shaped weight matrix is fitting, not plasticity. Note also that the connectome edge list and checkpoints are not in the repo (too large, no hosted download yet), so the wiring itself cannot be verified from a fresh clone.

Record

CategoryBrain models & bodies
DatasetsFlyWire, MuJoCo
LevelL1 — reviewed
Listed intownie/awesome-fruit-fly
First indexed2026-09-15

Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.