# fruitFly A whole fruit fly nervous system running as a spiking network on the GPU, a fly body it drives, and a bench for experiments on it. Install this package and a world holds the complete adult male *Drosophila* central nervous system (MaleCNS v1.0: 164,587 neurons, 25.6 million connections, CC-BY 4.0) as a leaky integrate-and-fire network you can stimulate, silence, read out and watch, plus a fly whose taste, looming, walking, turning, feeding, flight and escape run through that wiring. ## Contents The network: - `maleCNS.connectome/` is MaleCNS v1.0 packed for the GPU (109 MB, the edge table in two parts). It is an instance of the `connectome` asset type, which lives at the world root (`/zero/source/connectome.assetType/`) and travels with the package as a dependency: any nervous system's wiring as one asset (`manifest.json`, `neurons.bin`, `offsets.bin`, `edges.bin` or its parts). `guides { path = "types/connectome" }` is the reference. - `lif_integrate`, `lif_propagate`, `lif_clear`, `lif_reset` (`.computeShader`) are the four kernels of the network: membrane step, event-driven spike propagation over the CSR edge table, step close, reset. - `brain.module/` is `Brain.load(connectomeRef, params)`, which binds a connectome to the kernels and returns the network: `step`, `update`, `select`, `stimulate`, `silence`, `watch`, `rates`, `activity`, `classRates`, `groupActivity`, `drives` and the cloud. - `FruitFlyBrain.component/` puts a network on an entity: it loads the connectome the `connectome` field names, steps it every frame, draws the cloud, and exposes `stimulate` / `silence` / `watch` / `rates` / `reset` (the first three are synced, so every peer's copy receives the same inputs). - `neuron.shader/`, `neuron.material/` and `cloud_gather.computeShader/` are the instanced neuron cloud, each neuron lit by its own activity. - `crackle.soundClip/` is the network's own sound: a sparse crackle that is silent when the network is silent and a hiss when it runs away. The fly: - `flyBody.module/` is the fly's body as an entity tree with a tripod gait, wings, proboscis and a hop. - `FruitFly.component/` is the embodied fly: it reads the world into the brain's sensory neurons and reads motor commands back out of named cell types. - `FoodPatch.component/` is a disc of food on the ground the fly can taste. - `wing.material/`, `eye.material/` and the `wingbeat`, `feed` and `jump` sound clips are what the fly looks and sounds like. The instrument: - `FruitFlyLab.component/` is the bench window: network state and knobs, live readouts with traces and rasters, hot classes, live drives, a neuron panel, a legend, and a selection line to stimulate, silence, restore, count or watch any set of neurons. - `selection.module/` is the selection grammar the bench and everything else parse with. - `protocol.module/` is the guide's experiments as records anything can run, and the verdict a run computes from its own numbers. - `FruitFlyExperiments.component/` is the experiments panel: those protocols with one press each, a run list, a compare overlay and a verdict line. - `FirstRun.component/` is the four-step card that walks a newcomer from a reset network to a proboscis coming out. - `lab.inputMap/` and `experiments.inputMap/` are the bench's and the panel's controls, each binding keyboard, gamepad and touch. - `experiments.guide/` is the bench manual and the five experiments. The lab around it: - `labFloor.material/` and `FruitFlyLook.component/` are the floor, the fog, the vignette and the lighting intent, held as data rather than as code. - `arenaRing.mesh/` is the ring on the floor the fly stays inside. ## Usage ```luau -- The brain: one entity, one network. The default connectome is maleCNS. local brainEnt = entity.spawn("Brain") brainEnt.position = { 0, 6, -8 } brainEnt.component.add("FruitFlyBrain", { gain = 0.6, timeScale = 1 }) -- An embodied fly driven by it, and something for it to taste. local fly = entity.spawn("Fly") fly.component.add("FruitFly", { brain = brainEnt, bodyLength = 0.6 }) local food = entity.spawn("Food") food.position = { 3, 0, 0 } food.component.add("FoodPatch", { radius = 0.6, sugar = 1 }) -- The bench, the experiments panel and the card. entity.spawn("Lab").component.add("FruitFlyLab", { brain = brainEnt, fly = fly }) entity.spawn("Experiments").component.add("FruitFlyExperiments", { brain = brainEnt, fly = fly }) entity.spawn("FirstRun").component.add("FirstRun", { brain = brainEnt, fly = fly }) -- Or from code: local comp = brainEnt.component.get("FruitFlyBrain") comp:watch("MN9", { types = { "MN9" } }) -- proboscis motor neurons comp:stimulate({ types = { "claw_tpGRN" } }, 100) -- tarsal taste pegs at 100 Hz print(comp:rates().MN9) -- Hz per neuron, updated every frame ``` A selection names neurons by the dataset's own vocabulary: `types` (exact cell-type names), `typePrefix`, `pattern` (a Lua pattern over the type name), `classes`, `superclasses`, `sides` (`L`/`R`/`M`), `excitatory`, `indices`, `limit`. `brain:info(i)` answers what any neuron is. ## The model Every neuron is a leaky integrate-and-fire unit with the constants Shiu et al. 2024 (Nature) fitted to the FlyWire brain: rest and reset at -52 mV, threshold -45 mV, membrane time constant 20 ms, synaptic time constant 5 ms, refractory period 2.2 ms, transmission delay 1.8 ms, 0.275 mV per synapse. A presynaptic neuron whose consensus transmitter is GABA, glutamate or histamine inhibits; every other transmitter excites. A connection carries its synapse count; connections under `minSynapses` (default 5) carry nothing. `gain` scales every synapse. MaleCNS traces about 2.5 times the synapses per connection FlyWire does, so at gain 1 the network runs away under any drive; at 0.6 (the default) driving the 50 tarsal taste-peg neurons at 100 Hz holds the proboscis motor neurons MN9 at about 115 Hz while the Kenyon cells stay under 0.1 Hz, and driving the 595 Johnston's organ neurons instead leaves MN9 at zero. At 0.5 the taste drive reaches nothing. The bench's gain slider is the experiment, and the experiments panel's "Gain recovery" runs all three in one press. Everything runs on the GPU: the state, the input ring, the spike list and the edge table. The CPU sees the per-group spike counters each frame and, on request, the per-neuron activity lane. Integer atomics make a step order-independent, so two peers stepping the same inputs read the same network. ## The fly `FruitFly` reads named cell types out of the brain and writes the world into others; every name is in `FruitFly.component/init.luau`: | command | cell types | |---|---| | walk | the leg motor neurons (`vnc_motor` types ending in ` MN`), per side | | turn | `DNa01`, `DNa02` left minus right | | proboscis | `MN9` | | flight | the wing power-muscle motor neurons (`DLMn`, `DVMn`) | | jump | a `DNp01` (giant fibre) burst | | sense | cell types | driven by | |---|---|---| | tarsal taste | `claw_tpGRN`, `dorsal_tpGRN`, per side | that side's front foot on a `FoodPatch` | | labellar taste | `LB*` labellar bristle GRNs | the extended proboscis on a `FoodPatch` | | looming, LPLC2 path | `LPLC2` | a player's avatar growing fast in front of the fly | | looming, LC4 path | `LC4` | the same, driven as a separate path so one can be silenced | The two looming paths are driven separately and weighted by how head-on the approach is, so an avatar walking up behind the fly does not loom and silencing one detector class is an experiment with an answer. The body moves kinematically at `walkSpeed` body lengths per second; the fly faces its entity's -Z. On a shared world the entity's owner writes the senses into the brain (a synced call) and every peer poses its own copy from its own network. ## Controls | Action | Keyboard | Gamepad | Touch | |---|---|---|---| | Bench: show or hide | L | Start | The "Lab" button | | Bench: move focus | Down and Up, Tab | D-pad down and up | Tap the control | | Bench: activate the focused control | Enter | North (Y) | Tap | | Bench: nudge the focused slider | Left and Right | D-pad left and right | Drag it | | Bench: identify a neuron | Click it in the cloud | West (X) | Tap it | | Experiments: show or hide | P | Select | The "Experiments" button | | Experiments: next protocol | N | Right stick press | Tap the protocol | | Experiments: previous protocol | B | None: the list wraps | Tap the protocol | | Experiments: run | R | Right trigger | RUN | | Experiments: compare | C | Left trigger | COMPARE | Both maps name their controls apart (`lab*` and `exp*`) and take disjoint keys and buttons, so the bench and the panel can be open together and neither answers for the other. ## Data and license MaleCNS v1.0: HHMI Janelia FlyEM, Google Research and the Cambridge connectomics group, released 8 June 2026 (Berg et al. 2026, Cell, doi:10.1016/j.cell.2026.08.015), CC-BY 4.0. The pack was written by a script that reads the public flat-connectome export (`connectome-weights ... traced-only`, body annotations, consensus neurotransmitters), assigns dense indices, sorts edges by presynaptic neuron and writes the four files `types/connectome` describes; the manifest carries each source file's SHA-256. ## Notes - Loading the network reads the 102 MB edge table into a GPU buffer once per `FruitFlyBrain`; the VFS copy is released afterwards. - The neuron cloud is one instanced draw of every neuron with a soma (139,741). Its transforms are static; the activity lanes are written by the integrate kernel. - `maxStepsPerFrame` bounds GPU time per frame; at 32 steps of 1 ms a 60 fps engine keeps up with real time. - A neuron reads out under at most three groups at once. The fly holds seven of its own and the protocols watch five more under the guide's names, so a bench watch of the same cell types under the same name shares the group rather than spending another slot.

# FruitFly An embodied fly driven by a `FruitFlyBrain`. Every frame it writes the world into the network's sensory neurons and reads a motor command out of named motor and descending cell types. Which side of the body a sense or a command belongs to is the dataset's answer, not this component's. MaleCNS v1.0 labels a side on the looming detectors and on the leg motor neurons, and labels the tarsal and labellar taste neurons "unknown", so looming is driven per eye and taste is not. | sense | cell types | driven by | |---|---|---| | tarsal taste | `claw_tpGRN`, `dorsal_tpGRN` (60, side unknown) | either front foot on a `FoodPatch` | | labellar taste | `LB*` (165, side unknown) | the extended proboscis on a `FoodPatch` | | looming, left eye | `LPLC2` side L (94), `LC4` side L (71) | anything growing in the left eye's field, the arena wall included | | looming, right eye | `LPLC2` side R (91), `LC4` side R (55) | the same for the right eye | The two looming classes are separate drives, so silencing `LPLC2` leaves `LC4` driven and the ablation has an answer. | command | cell types | |---|---| | walk | leg motor neurons (`vnc_motor` types ending in ` MN`), left 189 and right 185, averaged | | turn | `DNa01` + `DNa02` left minus right, plus the leg motor neurons right minus left | | proboscis | `MN9` | | flight | `DLMn` + `DVMn` wing power-muscle motor neurons | | jump | a `DNp01` burst above `escapeHz` | Beating wings lift the fly only once it is already off the ground: on six legs the beat is a beat, and the legs keep walking underneath it. ## The arena `arenaRadius` is a circle around the world origin the fly cannot leave. Inside `arenaMargin` of it the wall is read by each eye as something looming, so the fly meets the wall with the circuits it already has; the body is clamped at the radius as well, so whatever the circuits do the fly is still in the arena. The wall is only a stimulus while the fly is carrying itself into it, and the detectors adapt to it over `WALL_ADAPT_S` and recover over `WALL_RECOVER_S`. A fly walking at the edge gets the burst that turns it; a fly resting against the edge is looking at something that has stopped arriving, and drives nothing. That matters beyond the fly: a standing drive on 165 looming neurons puts the whole network into a spreading state, and every reading anyone takes in this world is taken from that network. `escapeHz` is a RISE, not a level. The reading is how far `DNp01` has gone above the rate it has been holding, averaged over `ESCAPE_BASELINE_S`. The network's own recurrence holds the giant fibre at a standing rate with nothing looming at all, and a threshold on the absolute rate reads that as a permanent emergency and hops the fly for ever. Fields: `brain`, `bodyLength`, `walkSpeed`, `turnRate`, `legTurnGain`, `motorFullHz`, `flightFullHz`, `tasteHz`, `loomHz`, `loomFov`, `escapeHz`, `arenaRadius`, `arenaMargin`, `groundY`, `wander`, `wanderTurn`. `wander` and `wanderTurn` are the walking and the turning this fly does that its connectome is not asking for, and `motor()` reports both apart from `brainWalk` and `turn`, which are the network's. Set them to zero for a fly that does only what the wiring says; leave them and a quiet network leaves a fly that explores its arena instead of walking one straight line to the edge. Methods: `motor()`, `senses()`, `readouts()`, `sensoryNeurons()`, `drives()`, `soundVolumes()`. Events: `onJump`, `onFeed` (once per meal). ```luau local fly = entity.spawn("Fly") fly.component.add("FruitFly", { brain = entity.find("Brain"), bodyLength = 0.6, wander = 0.15, arenaRadius = 12 }) ``` `motor()` reports `brainWalk` (the leg-motor term alone) and `wander` (the baseline added to it) beside `speed`. `senses()` reports `sugarFootL` and `sugarFootR` per front foot, `loomLPLC2` and `loomLC4` in Hz, `loomSide` (-1 the left eye to +1 the right) and `wall` (0 to 1). The fly's own drives are filed under the source `fly`, so clearing them leaves the bench's and a protocol's standing. On a shared world the entity's owner writes the senses into the brain (a synced call); every peer poses its own copy from its own network.

# selection.module The grammar the lab bench, the experiments panel and the guide all read when a person names a set of neurons. A selection is a line of space separated `key:value` terms, and it turns into the table `brain:select` takes. | Term | Selection key | Shape | |---|---|---| | `type:MN9,DNa01` | `types` | comma list, merged across repeated terms | | `prefix:JO-` | `typePrefix` | one value, last term wins | | `pattern:^LB%d` | `pattern` | one value, a Luau pattern over the type name | | `class:Kenyon_Cell,MBON` | `classes` | comma list, merged | | `super:descending_neuron` | `superclasses` | comma list, merged | | `side:L` | `sides` | comma list, merged | | `excitatory:true` | `excitatory` | true or false | | `indices:0,1,2` | `indices` | comma list of neuron indices | | `limit:100` | `limit` | a number | `class:`, `classes:`, `super:`, `superclass:`, `superclasses:`, `side:`, `sides:`, `type:`, `types:`, `index:` and `indices:` are all accepted spellings of the same keys. Every key given must hold, so terms of different keys intersect and terms of the same key union: `type:MN9 type:DNa01` names four neurons, `class:Kenyon_Cell side:L` names the left Kenyon cells. ## Reading it ```lua local Selection = require("~.selection") local sel, err = Selection.parse("type:MN9,DNa01 side:L") if sel == nil then print(err) -- names the term that failed else print(Selection.describe(sel)) -- "types MN9, DNa01; sides L" print(#brain:select(sel)) end local brainComp = Selection.resolve(public.brain) -- the FruitFlyBrain, or nil ``` `M.PLACEHOLDER` is the field placeholder that names every term, and `M.KEYS` is the same list as one line for an error message, so the field, the error and this table cannot drift apart.

# brain A whole connectome running as a leaky integrate-and-fire network on the GPU. `Brain.load` binds any `connectome` asset to the package's kernels and returns a network; the `FruitFlyBrain` component is the usual owner, and anything holding the network can step, drive, cut, read out and draw it. ## Exports - `M.load(connectomeRef, params?) -> Brain`: allocate the network at rest. `params` overrides any key of `Params`: `dtMs`, `vRest`, `vThresh`, `vReset`, `tauM`, `tauSyn`, `refracMs`, `delayMs`, `weightMv`, `gain`, `minSynapses`, `traceMs`, `maxSpikes`, `seed`, `timeScale`, `maxStepsPerFrame`. A key that is not one of those raises, as `setParams` does, and so does a value that is not a number. - `brain:update(dt)`: once a frame: queue the steps `dt` covers, issue and drain the counter readback, and fold the activity snapshot when something has asked for one. - `brain:step(n)`: queue `n` steps outright. - `brain:select(sel) -> { index }`: the neurons a selection names, ascending. The array is the network's own and memoized on the selection, so read it rather than writing into it. - `brain:stimulate(sel, hz, source?)`: Poisson drive. The rate a neuron sees is the largest any source asked for, so the fly's senses and the bench's experiments can drive the same neurons without taking each other's drive away. `brain:drives()` lists `{ source, count, hz }` per source and `brain:clearStimulation(source?)` clears one source or all of them. - `brain:silence(sel, on)`: ablate or restore. - `brain:watch(name, sel)` / `brain:unwatch(name)` / `brain:rates()`: read groups out as Hz per neuron. A neuron carries at most three groups at once and a watch that would need a fourth on any neuron is refused, naming the neuron and the groups it holds. Watching a name again with the same selection is a no-op; `unwatch` frees the slots and the group's id. - `brain:classRates(limit?)`: the cell classes with the most activity right now, from the 60 ms spike trace. `brain:groupActivity(name)`: that trace per neuron for a watched group of at most 32, which is what a raster draws a row from. Both arm a per-neuron readback that is taken at most every twelve frames and only while something is asking, because it reads the whole activity lane. - `brain:activity()`: network mean rate, total spikes, dropped spikes, simulated time. - `brain:info(i)`: what one neuron is. - `brain:buildCloud(opts, onReady?)` / `brain:setCloudDrawn(on)` / `brain:cloudDrawn()` / `brain:cloudBounds()` / `brain:setCloudOrigin(xyz)` / `brain:showCloud(opts)` / `brain:hideCloud()` / `brain:worldPosition(i)` / `brain:neuronAt(screenX, screenY, radiusPx?)`: the instanced neuron cloud. - `brain:setParams(patch)`, `brain:setRunning(on)`, `brain:setTimeScale(x)`, `brain:setMaxStepsPerFrame(n)`, `brain:reset()`, `brain:ready()`, `brain:destroy()`, `brain:describe()`. ## Usage ```luau local Brain = require("~.brain") -- from inside the package local brain = Brain.load(asset.resolve("fruitFly.maleCNS", "connectome"), { gain = 0.6 }) brain:watch("MN9", { types = { "MN9" } }) brain:stimulate({ types = { "claw_tpGRN" } }, 100, "bench") -- every frame: brain:update(dt) print(brain:rates().MN9) ``` A selection is `{ types?, typePrefix?, pattern?, classes?, superclasses?, sides?, excitatory?, indices?, limit? }`; every given key must hold. ## The cloud, and why it is built in two halves A registered instanced population has its whole transform run recomposed on the CPU every frame, whether or not the cull draws any of it, so 139,741 neurons cost about 4.7 ms of frame whether or not anyone is looking at the brain. `buildCloud` fills the buffers and `setCloudDrawn` registers or drops the draw without touching them, which is what lets `FruitFlyBrain` put the registration down while the camera is pointed somewhere else and put it back the frame the cloud enters view. Taking it down leaves `worldPosition`, `neuronAt` and `cloudBounds` answering as before. Only neurons that have a soma get a slot, which is 139,741 of 164,587. Slot and neuron index are therefore different numbers, so the `cloud_gather` kernel copies each slot's activity out of the neuron it draws once a frame, and `slotToNeuron` maps back for picking. The fill is spread over frames from a task, about forty scan slices and eighteen upload slices, so building the cloud costs about a millisecond a frame for a second rather than one frame of a third of a second. ## Notes - The network is silent until something is driven; a spike always traces back to a stimulus. - Every GPU buffer is named `brain.<id>.<name>`; `compute.observe()` lists them. `destroy` releases them all. - The kernels' `params` are per shader, so two networks on one engine share the model constants the last `applyParams` wrote. - Editing this file hot-reloads its methods onto networks that are already running. A network records the `SHAPE` it was made with and `FruitFlyBrain` loads a fresh one when the module's shape has moved past it, so raise `SHAPE` whenever a field is added to or taken off a network or its cloud.