Overview

Two nested loops:

1. Outer loop, MAP-Elites over behaviour. The archive is a pyribs CVTArchive with about 500 cells over four descriptors measured in the nominal flat-floor episode: duty (fraction of time the ground-touching tips are on the floor), height (mean torso height over the body reach), legs (limbs on the torso with a joint and no wheel, so a hexapod never competes with a two-legged crawler for a cell) and cot, the cost of transport (mechanical work per weight and distance, on a log scale). The last one lets an efficient walker keep its own cell next to a faster but hungrier roller. Each cell keeps the best body found there, so a run ends with a map of specialists instead of a single design. Parents are drawn with probability proportional to their novelty (mean distance to the 10 nearest elites in descriptor space), which pushes the search towards sparse regions.
2. Inner loop, CMA-ES over the controller. For a fixed body, CMA-ES searches the controller genes that this body actually uses. A promising body is therefore judged by the best gait it can learn, not by a lucky random controller.

Children are mutated copies of elites, or grafts of two elites. The child starts its inner loop from the parent controller, so learned gaits are inherited along with the body (Lamarckian inheritance). Controller genes are indexed by the stable slot id of each module, so a mutation that keeps a module keeps its control. Only the genes that affect the child are optimised; the rest stay in the genome and become active again if a module returns.

Construction rules act as natural selection by the printer. A body is buildable only if every part fits a 220 mm print bed, every link bends under the 2.94 N*m servo stall torque with a stress of at most 25 MPa (half of what PLA takes), and no parts overlap at spawn. The torso is not limited: its outline is split into bed-sized tiles joined by splice plates, and each spine segment is its own plate. A body that breaks a rule scores as unbuildable and is never simulated.