Published neural foundation
The complete Shiu/FlyWire v630 network: 127,400 neurons and 14,687,178 weighted connections. Our earlier run produced the recorded MN9 spikes and matched the original code for one full trial.
Digital fly / Experiment 02
First movement.
A signal acquires a physical consequence.
The same brain recordings from First taste, now connected to a 3D NeuroMechFly body. Motor spikes apply a turning force to its mouth joint. A physics simulator computes what moves.
One real step, one explicit assumption. The body comes from a scientific model. The spike-to-force bridge and mouth-joint settings are our engineered approximation—not a validated simulation of the fly’s feeding muscles.
Ready · recorded physics
Joint displacement Individual motor spikes
132 motor spikes in this trial— maximum displacement
Both the brain run and the physics run happened locally before this page opened. You are inspecting recorded physical states, not running a live brain or physics engine in the browser.
Two extra causal checks
Where does the movement come from?
Both checks reuse the very same sugar spikes from trial 1. Only the physical coupling changes.
Measured physical consequence
One bridge. Different neural outputs.
Maximum angular displacement from the settled pose. Each number is one physics replay of a verified neural trial. These angles describe this engineered apparatus, not measured living-fly behavior.
What was inherited, and what did we add?
Scientific body geometry
NeuroMechFly’s micro-CT-derived meshes, segment masses, supplied joint anchors and neutral posture. Rostrum and haustellum are separate body segments. The rest of the body is held still in this experiment.
Our engineered bridge
Each motor spike adds a decaying activation signal. A fixed saturating rule turns that signal into torque. The bridge receives only spike times—never “sugar” or “bitter.” All trials use the same parameters.
Our engineered mechanics
One sagittal hinge at the supplied rostrum anchor, with a spring, damping and a torque motor. MuJoCo solves its motion using the body inertia. The haustellum is fixed to the rostrum; we have not recreated separate labellar joints, muscle 9, or the natural multi-joint coupling.
What makes this physics instead of a condition-triggered animation?
The offline runner writes torque to an actuator and advances MuJoCo at 0.1 ms steps. It never sets a mouth angle while the simulation is running. The stored angles and body transforms are simulator outputs. The viewer displays those transforms.
“Wires off” eliminates the neural output. “Actuator off” preserves that output but blocks its mechanical effect. “Stiffer joint” preserves spikes and torque yet changes the motion. Halving the physics timestep changes the sugar-trial trajectory by less than 0.01°.
What is still missing?
No simulated sugar contact, swallowing, walking, sensory feedback, or closed loop. The taste condition still selects an earlier neural experiment. The body is tethered; the plane is a visual reference. The force mapping, spring, damping and range are engineering choices. Before each run, the passive joint settles under gravity; that starting posture is not a claim about a living fly’s resting pose. MN9 is biologically associated with rostrum protraction, but a quantitative spike-to-muscle-force relationship has not been established here.
This demonstrates an open-loop connection between two scientific models through a disclosed artificial actuator. It is a first physical bridge, not a validated model of natural feeding.
Sources, measurements and reproduction
Shiu et al. · neural model · McKellar et al. · proboscis motor anatomy · NeuroMechFly v2 · Nature Methods · Pinned FlyGym source
FlyGym 2.1.0 · MuJoCo 3.9.0 · 17 body trials plus a timestep check. Model, recordings, hashes, assumptions, checks and reproduction commands are preserved in the experiment guide.