Recorded model run · Choose a camera or pause to inspectRecorded model run · Three camera views
A fly with a shotgun faces approaching spiders on a kitchen table. This is a recorded model run. Choose Chase, First person or Table to follow the same attempt from three angles.
The model receives two 64 × 64 RGB images and its own pitch, ammunition, reload state and previous movement commands. A learned visual encoder feeds a rate network containing all 166,700 retained MaleCNS cells and 25,582,938 directed connections. Both the movement and firing outputs and the spider segmentation maps are read from that network. No enemy coordinates, color threshold detector or teacher actions enter the recorded inference. Detection and action are parallel outputs: the displayed boxes do not steer the fly.
Training uses supervised demonstrations for actions and rendered, depth-tested spider geometry for detection labels. The visual encoder, sensory projection, gains and biases across the full network, and output heads are trainable. Contact topology and approximate neurotransmitter signs stay fixed. This is imitation learning, not learning from rewards. Each observation starts a fresh state calculation with three graph updates. Neural state is not carried between observations; previous movement commands are provided as inputs.
The eye panels re-render the recorded viewpoints. Frames mark connected regions in the learned spider masks for each eye. The brain panel displays 6,000 positioned cells sampled for drawing from the full calculation. Their recorded signed activations drive the flashes; only drawing is sampled. Brightness shows deviations from each cell's recorded mean, scaled by its variation. These are model activations, not measured spikes. Link brightness follows its endpoints, not recorded synaptic current. Colors identify anatomical groups, not proven task roles.
An audit of all 900 decisions found that visual sensory and optic-lobe cells dominate the action readout. Replacing all VNC final states with their recording means changed each movement output by less than 0.00002 and changed no fire decisions. This tests readout dependence on this run, not biological function. VNC flashes should not be interpreted as control of the legs.
This is an engineered model built on the full connectome. It does not reconstruct biological compound-eye optics, measured neuron dynamics or muscle-level control. Walking legs and the weapon follow model commands through animation. Spiders use a scripted chase controller; physics, cooldowns and movement smoothing are part of the environment. Reloading is automatic. Shot collision uses simplified hitboxes and an eye-centered aim ray; the offset weapon is animated. The two cameras are synthetic stereo vision, not a reconstruction of a fly's compound eyes.