DROSOPHILA AEROSPACE

Mission control.

Connecting the pilotFSP / 009
Flight options
FLIGHT FEED
DISPLAY SYNCFLIGHT 001
ATLANTIC RECOVERY ZONE
LANDING SCHOOL

Bring it home, little guy.

Vertical descent · calm air

FLIGHT TIME00:00.0
🪰
PILOT IN COMMANDFly / Eyes, body and antennae
Loading the complete flight network…
Waiting for first touchdown
ALTITUDEm
VERTICAL SPEEDm/s
ATTITUDE°
PROPELLANT%
DECK ERROR — mLATERAL SPEED — m/sSTYLE BONUS PENDING
THE EXPERIMENT

Small brain.
Ambitious career change.

The network includes 166,700 annotated neurons, 25,582,938 directed connections, and 124,177,617 synaptic contacts from MaleCNS v1.0. Every connection between retained, annotated non-glial neurons is included, with no weight cutoff.

The brain’s shape comes from 3,487,943 measured synaptic endpoints and soma fallbacks. Drawing samples one in 128 endpoints plus a measured anchor per located neuron; graph computation keeps all retained edges. This is a point cloud of measured anatomy, not complete neuron skeletons. 121 neurons have no available position and participate in the network without a drawn location. Brain view crops the nerve cord; Whole CNS reveals it.

The complete controller

Every control decision evaluates all 166,700 retained neurons and all 25,582,938 directed connections twice. For the perceiving fly, 1,536 light pixels and 22 body/odor signals enter sensory cells selected by their actual annotations. Photoreceptor positions use synaptic links to annotated optic columns, with documented nearest-column fallbacks. Body response directions and optical calibration are modeled. A trained readout uses all 1,314 descending neurons and 815 annotated motor neurons to generate ten commands. No separate small pilot supplies or overrides activity. The flight waits for a decision, then advances three 50 ms physics steps near the surface or two 250 ms steps aloft during orbital missions; playback speed changes waiting time, never the calculation. The atlas shows that same computed activity.

What actually learns?

The external output decoder learns; the anatomical graph’s individual synapse counts and signs stay fixed. The perceiving fly learns independently of the instrument-trained pilot. The landing curriculum first fits full-network activity to the brightness of presented altitude and vertical-speed indicators, together with body feedback. The same indicators are visibly rendered on the cockpit monitor and reach the fly through its modeled eyes. Flight reward then trains throttle. The resulting coefficients are folded into the ordinary 21,300-weight readout. There is no flight teacher or hidden navigation input. The first lesson keeps pitch, roll, yaw and gaze neutral and uses the stationary Landing school deck. Local reward training tests paired weight changes in complete graph-driven flights with identical starting seeds, keeping more safe landings first and higher reward second. Every candidate advances the full graph at every decision. Training runs on your device, takes longer than the earlier small circuit, and saves local checkpoints.

Simulated arousal

The optional neural-state experiment multiplies each neuron’s activation input by 1.00, 1.02 or 1.05 before its tanh response. Every retained neuron and connection still participates. This is a global sensitivity experiment inspired by fly neuromodulation research; it does not simulate a chemical dose, receptor dynamics or the cell-specific effects of octopamine. Training keeps that gain fixed while learning output weights, and saves it with the checkpoint. Exact decision replays include the gain. The paired experiment changes gain while holding weights and initial conditions fixed; it tests operation, not whether stimulation improves learning speed.

Dopamine can regulate synaptic plasticity in fly learning circuits, but this app’s reward search does not model that biological mechanism. External ethyl acetate and geosmin stimuli are modeled separately through DM1 and DA2 olfactory cells. They are odors, not pheromones or internal neuromodulators. Automatic release depends on sampled retinal light/change and body rotation, not hidden navigation targets.

Recovery, with flair

Fly with flair adds a full yaw-turn target. A turn earns up to 25 extra points only after completing roughly 360°, settling rotation and attitude above the deck, and landing safely. Low, excessive or badly tilted turns lose eligibility. Repeated spins and oscillations do not multiply the bonus. Crashes earn zero style points. The live label separates a pending maneuver, recovered turn, and banked landing bonus. Turn flair off for a recovery-only assignment.

Activity is a simplified signed rate model, not recorded biological firing or a validated emulation of a living fly. The sensory mapping, readout, signs and learning rule are modeling choices. There is no language model flying the rocket. The orbital program adds launch, circular insertion, at least one full revolution, deorbit and barge recovery on a world with a 6 km radius. Only the instrument-trained reference receives orbital guidance cues. The perceiving fly must learn from its sensory observations; no successful orbital skill has been demonstrated for it. This is a compressed orbital sandbox, not an Earth-scale mission or a learned mission planner. The perceiving fly samples monitor light through two modeled eyes. The reference pilot samples numerical instrument readings. The ten outputs move controls with limited travel speed; inverse kinematics places limbs on those controls. Facial expressions and radio jokes are cosmetic.

The fictional vehicle

This stylized 3D booster has a center engine and a selectable three-engine bank, four grid fins, attitude jets, and a tiny Dragon-style crew pod with an antenna. Thrust, aerodynamics, faults and landing tolerances are game-scale approximations. It is not a flight-qualified Falcon 9 model. Twenty-seven profiles cover moving deck patterns, wind shear, precision approaches, fuel limits, slow controls, spinning entries and combined faults. Engine-out profiles completely shut off the center engine. Night and fog also affect the perceiving fly’s rendered camera light.

A fluid flight feed

The display follows your screen’s refresh rate, with interpolated rocket and limb motion between physics steps. Physics runs at 20 Hz near the surface and 4 Hz aloft in orbital missions; every controller decision still runs the complete graph. Rendering can briefly trail the exact telemetry; pausing aligns the pose with the current actuator state. Frame rate depends on the device and browser.

Inspect the anatomical sensory map ↗ · Visual learning and ablation results ↗ · Orbital stability regression ↗

Follow a connection

Choose a control to follow measured incoming connections through four influential output neurons, the learned readout and the highlighted limb. The full readout still includes all 2,129 output cells. Atlas lines join measured neuron anchors; they are not reconstructed axon paths.

Try an experiment

Click a neuron to inspect its source ID and annotation, then pulse it to inject eight network updates of excitation. Spotlight the descending and motor outputs to see where commands leave the brain. Throw a gust to challenge recovery, or take the controls yourself.

The full-network evaluation is loading. Its report includes every seed and failure.

The complete anatomy and graph download about 91 MB on first use. Flight begins only after the complete graph is ready. A current desktop browser with WebGL 2 works best.

Fly Space Program is an independent, experimental research project. Source and documentation ↗ · No Theo License ↗. Source-available; not open source.