Haltere v0.1.0: the fly brain flies in Liftoff

First release: a connectome-constrained fly brain (Janelia male CNS v1.0, 30,000 neurons, 2.77 M synaptic connections) that flies an FPV drone in Liftoff.

Trained brains (slim inference checkpoints, 12 MB, are also in the repo under artifacts/; the files here include the optimizer state for resuming training):

• ftSmooth_best.pt – recommended for Liftoff: fine-tuned with 50 ms extra latency and a smoothness penalty. Simulator: 0.30 m mean error, 95% within 0.5 m, no crashes at full difficulty.
• ftRobust_best.pt – wide domain randomization; first brain to fly in Liftoff (2 m hover, 0.34 m mean error over 40 s).
• imJ_best.pt – imitation of the MLP baseline with the premotor readout; 0.22 m / 95% in the simulator.
• mlp_baseline.pt – the MLP controller used as the imitation teacher.
• flight.npz + flight.nodes.parquet + flight.meta.json – the built flight graph (neurons, signed synapse counts, populations); rebuildable with haltere build from the public connectome files.

Videos: liftoff_hover.mp4, liftoff_square.mp4 (in Liftoff, with live neural activity), flight.mp4 (simulator).

Everything else (the 570 MB raw connectome) is downloaded by haltere fetch from the Janelia FlyEM public bucket (CC BY 4.0).