Repo

fly-self-driving

Steers a car from 64×32 pixels with a fixed fly connectome by learning synapse gains and neuron leaks, while keeping recurrent state across decisions.

GitHub ↗Sep 13, 2026Paper ↗

  • robotics
  • imitation-learning
  • connectome
  • dagger
Pipeline diagram showing method flow with feedback loop and processing stages
Pipeline diagram showing method flow with feedback loop and processing stages

What it is

Trains a fixed fruit-fly connectome to steer a simulated car from pixels, reaching 58/60 held-out traffic streets with no car collisions. Drive a simulated street using a measured 165k-neuron fruit-fly connectome with fixed wiring, trained by behavior cloning plus DAgger and persistent recurrent state.

Key results

  • 58/60 held-out traffic streets completed, with no collisions with any car
  • Persistent state raised first task success from 3/20 roads to 17/20
  • Trains 25,728,319 constrained params in about 35 min on one H100

Shows a measured insect connectome can be trained as a large fixed-topology controller for closed-loop driving from pixels.

One useful ML email a week.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.