§ MOTOR MAPPING
From grid state to torque: which number moves which leg. Trace a cell to the joint it drives — or pin one and override it on the live fly.
The evolved rule produces an 8×8×4 grid of numbers every tick. To drive the body we read channel 0 of the top-left 7×6 block, row by row — 42 numbers, one per joint. Number i becomes the target for actuator i, rescaled clip(u,−1,1)·3.14 rad: the rule's unitless −1…1 output turned into a joint angle of ±180°, which the physics engine then drives that joint toward. Hover a grid cell or a leg joint to trace that wiring — the grid position is deliberately unrelated to which leg moves.
s₀[0,0]→u[0]→joint—q*[0] = clip(u,−1,1)·3.14 radOverride one joint
Pin a cell, then this slider overridesit — forcing that one joint's target to a value you pick instead of letting the evolved rule set it. The other 41 keep walking, so you can see exactly which leg that single number drives. The slider is the raw output u ∈ [−1, 1]; the fly receives clip(u,−1,1)·3.14 rad.
pinned i=0 · LF · coxa · yaw
Clamps that one joint's target; the other 41 keep running the evolved gait — live MuJoCo.
The alternatives are the interesting part
One cell → one joint is just the cheap wiring. The directions worth building out — and worth visualizing — are richer: a learned readout that lets evolution place the motor taps itself; a population code where many cells vote on each joint instead of one cell owning it; and the real biological version, a descending-neuron interface— command neurons projecting from the brain onto the leg controllers. That last one isn't hypothetical here: the Giant Fiber (DNp01) firing rate is already read out as the escape drive in the embodied loop. See the descending interface on the Embodied tab →
What we chose
Channel 0 of the top-left 7×6 sub-grid, read row by row, gives 42 numbers; entry i becomes the target for actuator i (manifest order), rescaled clip(u,−1,1)·3.14 rad. One CA tick per control step. (Full equation in the appendix.)
Why
A wiring convenience, nothing more. There is nothing biological about this particular cell-to-leg assignment — a fixed sub-grid read in a fixed order was simply the cheapest way to turn a grid into 42 joint commands. The decorrelation you can see between grid position and which leg moves is the tell.
Alternatives
A small learned readout layer (let evolution place the motor taps); a different sub-grid, shape, or topology; or a population code where many cells vote on each joint instead of one cell per actuator.
The biological / frontier version
In a real animal this is the descending-neuron readout: command neurons projecting from the brain onto lower-level controllers. That interface — the hard part Eon names — is now built: in the embodied loop the Giant Fiber's (DNp01) firing rate becomes the escape drive. It is still a hand-tuned, single-neuron stand-in for the whole descending hierarchy, but it is the realized version, not a someday. See the Embodied tab.
The math behind this — equations and constants — is in the appendix.