FLY ATLAS
MaleCNS v1.0Let's be in touch

FlyTilt: learning joystick control

LEARNING THE CONTROLSLOADING RECORDINGS
01 / FIRST MOVEMENTSSTUDENT RUN
WAITING FOR DATA00:00 / 00:00
Preparing the experimentLoading recorded poses and neural activity
00:00
RECORDED MODEL ACTIVITY · SIMULATED HARDWARE

Learning the joystick

The full MaleCNS graph supplies activity to a readout of 67 right foreleg motor cells. The current experiment starts from previously trained controllers and uses reward search to adjust joystick gains and response time. A speed limit keeps joystick motion gradual.

The recording includes selected training attempts and their reward history. Lit checkpoints specify the route, so the model learns control and braking rather than independent maze planning. Successful takes retain the full attempt; practice takes may be short excerpts.

The state readout was calibrated on observed states from random interactions. The current reward search uses no teacher actions. Smoothing affects the controller and physics during training, as well as the recorded motion.

The body, sensory encoding, neural dynamics and servos are engineering approximations. The fly animation maps motor commands to a joystick grip; it is not a reconstructed muscle simulation. Neural flashes show modeled activity changes, not measured biological spikes.