§ BODY
The thing being controlled: a real FlyGym Drosophila in MuJoCo physics, running live in your browser — the evolved NCA walking it.
Live MuJoCo-WASM — short rollout of the real evolved controller, not a recording. The schematic below is the static plant it drives.
The same body, drawn as the plant the controller has to move. Below: a top-down selector of the six legs over the 7-DoF kinematic chain of the one you pick — and the exact slice of the 42-vector u it drives. Watch it walk above; trace which u[i…j] moves which leg here. Hover, tap, or focus a leg.
What we chose
NeuroMechFly / FlyGym in MuJoCo — the only biologically-grounded part of the whole system (real morphology, masses, contacts, from an X-ray microCT scan). A Drosophila with 42 leg position actuators (kp = 50) across six legs — 7 DoF each — ~87 joints, contact-rich. Control runs at 250 Hz (dt = 0.004 s) with 40 physics substeps at 10 kHz per control step; units are mm.
Why
The body is the testbed: the fixed substrate every controller is judged in. Nothing about it changes between experiments — swap the controller, the sensing, the objective, the optimizer, and this same plant is what has to actually walk. Make it real and the comparisons mean something.
Alternatives
Vaxenburg et al.’s whole-body RL fly (a learned policy, not an explicit plant to reuse); generic hexapod models (cheaper, but not a real animal’s morphology); kinematic replay (joint angles played back with no controller in the loop — looks alive, controls nothing).
The biological / frontier version
The full sensorimotor body — vision, olfaction, leg adhesion, muscle-level actuation — i.e. NeuroMechFly v2’s broader feature set. Today this plant is position-actuated and walks blind; the frontier closes the perceptual and muscular loops around the same morphology.
The math behind this — equations and constants — is in the appendix.