Video

A fruit fly’s brain, driving a street.

165,122 neurons and 25.56 million measured synapses, steering from a 64 by 32 pixel windscreen, on two streets it never trained on.

Traffic, parked cars and an overtake

The best model on the traffic task: a fly connectome with carried state, a 16-decision training window and three rounds of DAgger. Two held-out streets back to back, 9000 and then 9006, the hardest of the twenty by lane excursion.

Replayed from recorded state · 50 s

Moments in this drive

58 of 60 streets completed with traffic across three independent street sets; both misses left the road, none hit a car.
Lane offset within 4 cm of the expert on every set.
The same recipe on a randomly rewired graph completed 16 of 20.

What you are looking at

Left, the replay: A cinematic rebuild of the recorded drive: the fly is the car, every other car and crosser follows its logged position. Nothing here is fed to the brain.
Top right, what the brain sees: The actual 64 by 32 grayscale frame the network consumed at that tick, quantised to 8 bits after the fact.
Right, the connectome: All 140,024 neurons with measured soma positions, each coloured by its own recorded rate state: orange positive, blue negative, grey quiet, scaled per neuron over the drive. Brain above, ventral nerve cord below.
Bottom strip: Steering, speed where the model controls it, distance and lane offset, read from the same log.

measured graph · state carried · window 16 · DAgger ×3