One small fly. One whole orbit.
- LaunchPending
- Reach spacePending
- Stable orbitPending
- One revolutionPending
- Deorbit burnPending
- Atmospheric entryPending
- Barge touchdownPending
A small Newtonian world keeps the full journey watchable. The revolution counter starts after stable orbit; a successful mission requires that full circuit and a safe return to the barge. Engineered trajectory cues guide the trained neural controller.
What actually reaches the fly.
Light and body signals are active.
Upper light: clearance from 0 to 150 m. Lower light: vertical speed from −25 to +25 m/s; half brightness means zero. These displayed measurements reach the fly through its eyes.
These images feed annotated photoreceptors. Each retinal pixel averages sixteen rays across its area, reducing flicker from thin markings. Head movement changes which monitor pixels reach each eye. The deterministic optical model shows ocean, deck and sky; it is a simplified renderer, not the spectator camera. The fly receives these rendered camera and instrument pixels. The landing indicators show measurements; they do not prescribe a control action.
A stimulus the antennae can receive.
Steady rotation with visible monitor light releases ethyl acetate. Rapid rotation or image change releases geosmin. Exposure fades over time. These are simulated external odors, routed to matching annotated receptor classes.
No releases yet
Relative exposure, not a molecular dose. Odor does not select an action or change the learning reward. This rate model has no established innate odor preference.
These instrument cards are mission-control context. Only eye pixels, modeled joint position, body loads, rotation, touch, vibration and receptor-specific odor enter the perceiving fly.
A little experience goes a long way.
Inside the flight circuit.
Flight school
CHECKPOINTEvaluating the first visual lesson.
Training rotates through different conditions and preserves a separate pilot for each engine family.
Loading the flight record.
Landing tests and sensory controls ↗ · Historical instrument-pilot flights ↗A little more alert.
Change the circuit’s responsiveness, then fly or train at that setting. More stimulation can help or destabilize a learned pilot.
Baseline circuit gain.
Identical weights and starting conditions, with only circuit gain changed.
Inspired by neuromodulation research. This global gain rule does not model octopamine concentrations, receptor biology, dopamine-gated plasticity or pheromone perception.
Read the fly-flight study ↗From wiring to movement.
Mint raises a signal; orange lowers it. Select a neuron to find it in the atlas. The matching limb lights up in the cockpit. Pause to align the animated pose with the recorded control state.
Waiting for a network-driven control sample.
Inspect the measured edges
These are the strongest final-pass inputs into four output neurons with large readout contributions. Removing one shown edge is an isolated replay of that last pass, with every other input and earlier state fixed. It is not the effect of removing that connection for an entire flight.
| Source → output neuron | Contacts | Modeled sign | Input term | Command change if removed |
|---|
Incoming connections are measured fly wiring. The output weights and control-to-limb linkage are engineered. Joint positions follow the controls; they are not learned biological muscle activations. Displayed values are rounded; the edge table includes full precision on hover.
What drives the controls.
The exact input used by the selected controller. The perceiving fly receives 1,536 eye pixels and 22 body/odor channels. Eye ablations replay a whole eye together; pixel values are available in the decision sample.
Waiting for the first control decision…
Mission-control position & drift
● Rocket ■ Deck center
Arrow = relative velocity × 2 s; not a predicted path.
Dashed circle = 11 m position limit.
Descent cue
Mission-control context. The perceiving fly does not receive this cue or numerical instrument memory. The reference pilot does.
Instrument memory
Requested → actual
Select a control to inspect its inputs. Filled bar = requested position; white mark = actual position now.
Strongest input effects on Throttle
Pause & inspect computes the largest changes when each encoded input is set to zero on its own. Positive raises the command; negative lowers it. Bars span −2 to +2 command units.
What changes the Throttle command?
Each row replays this decision with only that encoded input set to zero. Everything else—including weights and the exact pre-decision recurrent state—stays fixed. The signed change is live command minus replay, on the −1 to +1 command scale. Positive means the live input raises this command; negative lowers it.
These isolated effects are not additive. They describe this calculation, not the fly’s intentions or a physically simulated intervention. Values are rounded for display; hover numbers for full precision.
| Input | Sampled value | Encoded input | Command change (−2 to +2) | Zero means |
|---|
Every input is clipped to [−3, +3] before the full network. Gimbal and jet bars are signed; throttle and selector bars start at zero. Fin bars show lever positions; a jammed fin can differ. Fuel exhaustion overrides throttle. The selector switches engine banks after its travel completes. Rendering effects never enter the sensors.