B-04 · CELLULAR GAITS
Cellular Gaits
A real fly connectome, run as a spiking brain, drives a real fly body — a looming threat becomes a Giant-Fiber spike becomes an escape bolt, closed through the physics. The evolved cellular automaton that walks that same body is the null-model rung this climbs from: the question throughout is whether structure, embodied, is enough to produce behavior.
§ THE FRAME
Three layers sit between wiring and behavior: connectome → dynamics → behavior. A connectome is a static graph; what an animal does is the dynamics that graph runs and the body those dynamics move. The gap between them is the whole problem.
Structure under-determines behavior. The same circuit can sit in an ordered regime or a chaotic one; the same motor map can produce a clean gait or a stagger. Knowing every synapse does not hand you the walk — you still have to specify the dynamics, the sensing, the mapping from neural state to torque, the objective being met, and the process that tuned it.
The body is the testbed. Here that body is a real FlyGym Drosophila — 42 leg actuators, contact-rich physics — and the controller is a neural cellular automaton: a single local rule run on a grid, the canonical toy model of emergence-from-local-rules. Each tab isolates one modeling choice — the body, the controller, its sensing, the motor mapping, and the search and its objective — and asks what was chosen, why, what the alternatives were, and where the biological version sits. The math is in the appendix.
§ ORIENTATION
The best evolved rollout, with the cellular-automaton state alongside it. Scrub the video; the four-channel grid on the right tracks the simulation tick by tick. The amber outlines mark the 42 cells wired to leg actuators. This is the object every other tab takes apart.
§ THE TABS
One modeling choice per tab. Each asks what was chosen, why, what the alternatives were, and where the biological version sits — then shows it running. All the math lives in the appendix.
- Bodythe plant we drive — a FlyGym Drosophila, 42 leg actuators across ~87 joints, walking live.
- Controllera neural cellular automaton parked just inside the edge of chaos — and the gain→gait sweep that shows why.
- Sensing & Loopopen-loop by default (the rule never reads the body) vs. the closed proprioceptive loop that now recovers from a shove.
- Mappinghow 42 grid cells wire to 42 joint targets — a convenience, not biology.
- Search & Objectivewhat fitness rewarded — forward distance, with a stability penalty that never fired — and the CMA-ES search that tuned 660 parameters from a stagger to a gait.
- Behaviorsclosed-loop sensorimotor behaviors — escape (the connectome bridge), perturbation, chemotaxis.
- Embodiedthe real connectome in the loop: a looming threat → the Giant Fiber → an escape bolt.
- Appendixthe math — update rule, motor mapping, fitness, CMA-ES, and the criticality instruments.