The next release must let one fly find and consume food and water placed away from its mouth, then search again when needed. Flight, landing and visible energy/water reserves belong to this delivery. Walking is an intermediate step. This is the canonical scope and acceptance plan; the linked guides describe what exists today.
| Capability | Status |
|---|---|
| Physical body and city | Native MuJoCo, rendered with Three.js / WebGL 2 |
| Taste → neural activity → mouth movement | Implemented, with causal comparisons |
| Finite portions, reserves, depletion and death | Implemented as simplified resource accounting |
| Distant sensing and neural need signals | Missing |
| Neural commands → learned locomotion → browser body | Missing |
| Search, flight, approach, landing and repeated feeding | Missing; delivery is incomplete |
Offer at mouth tests contact feeding. An apple placed nearby produces no attraction. A browser reproduction on 2026-09-13 observed 25.12 seconds of physical time with a noncontacting apple: no changed body pose, new motor trial or intake. Passing contact tests did not cover seeking.
The recorded flights use a learned policy and geometric navigator in a separate offline pipeline. An exploratory walking-policy port has produced isolated physical movement; numerical agreement and browser integration remain unverified. Neither establishes browser locomotion.
The fly starts with reserves, explores when it needs resources, approaches a detected source, lands and consumes it. It can rest when supplied and search again when needed. Uninterrupted flight is not the acceptance condition.
flowchart LR
Sources["Food / water<br/>fields"] --> Sensors["Local senses<br/>and needs"]
Sensors --> Brain["FlyWire<br/>commands"]
Brain --> Policy["Learned<br/>motor control"]
Policy --> Body["Physical motion<br/>and feeding"]
Brain -->|MN9| Body
Body --> Intake["Finite intake<br/>and reserves"]
Body --> Sensors
Intake --> Sources
Intake --> Sensors
This is the target architecture, not the current runtime. Field equations, need modulation and encoding/decoding gains are engineering approximations. Apple uses an odor cue; water needs a separate moisture cue or explicitly artificial channel. A visible halo may illustrate a field; it does not steer. An exhausted source stops emitting and cannot replenish either reserve.
- Hybrid control: the connectome supplies behavioral commands; validated flybody policies coordinate legs/wings. Policies and command decoders receive no food coordinates, object IDs or precomputed target direction.
- Physical movement: native actuators produce motion. Preserve collisions and aerodynamics; no pose/velocity writes, interpolated flight or episode resets during a run. Cruise flight does not prove takeoff or landing.
- Same brain: keep FlyWire 630 and all stored connections. Pin the provenance of new sensory/readout IDs. Preserve neural state across controller updates.
- Same renderer: keep Three.js / WebGL 2. The current physics and NumPy neural core run on a Docker CPU backend, not WASM/WebGPU. Consider a WASM solver only against measured need and numerical comparisons on this graph.
- Separate protocols: retain the bounded antennal/taste assays as regression tools. Their rest-start and rostrum-only rules apply to those experiments; locomotion needs its own clock, actuator allowlist and cancellation contract.
The runtime may enforce physical transition guards, resource limits and death. It must not select a source or inject steering behind a neural display. Saturated activity or unusable directional readouts fail the neural gate; do not hide them with pruning, repeated state resets or a scripted pilot.
Complete the first failing gate before unrelated work. All four are pending. Update status only with implementation and measured evidence.
| Gate | Work | Required evidence |
|---|---|---|
| 1. Continuous sensory decisions | Pin bilateral sensory/readout mappings, encode separate deficits and retain neural state | Left/right responses, changes after source relocation/removal, no-input and blocked-path comparisons, bounded activity and measured runtime |
| 2. Physical locomotion | Validate official walking inference, connect neural commands, then integrate flight and transitions | Forward/left/right/stop through actuators; blocked commands remove commanded motion; physical takeoff, sustained flight and landing |
| 3. Find and consume | Connect fields, locomotion, mouth contact and reserves in the browser | Distant apple and water each replenish the correct reserve; exhausted sources disappear; another source is found |
| 4. Complete release | Run actual UI, causal controls and sustained workload checks | Every acceptance case below passes on the identified commit, documentation matches and CI is green |
Gate 1 failed the first sensory feasibility measurements (2026-09-13). The original full-graph odor response persisted after input ended. An explicitly separate adaptive-threshold experiment recovered rest in six repeats but lost the response to the second cue and supplied no usable direction. See the protocol, numerical comparisons and results. The reference core and live assays remain unchanged.
Next implementation task: resolve the sensory path and bilateral readout mapping, including responsiveness to a second cue. Adaptation alone did not pass this gate; it is not ready for motor integration. The isolated walking probe starts gate 2; the recorded flight controller supplies a cruise reference, not a validated transition system.
The navis, hae and TMNF-C sources were inspected on 2026-09-13. None is integrated into the project; their reported learning and performance results remain unverified locally. These investigations support the existing gates:
The connectome data inventory records public download sources, verified offline annotations and candidate ID correspondences for this work. FAFB releases, BANC and MaleCNS remain separate; no runtime data was migrated.
| Investigation | Place in the plan | Evidence needed before adoption |
|---|---|---|
| Anatomy with navis | First: resolve gate 1's sensory and steering candidates | A small, sourced table of FlyWire 630 IDs, sides, cell types and connections; morphology comparisons with explicit units and brain spaces |
| Continuous directional response | Next in gate 1, using the checked mappings | Repeated left/right cues and removal without resetting neural state; useful readouts on the second cue; no-input and blocked-path controls |
| Event-driven C/WASM core | Address the measured neural runtime shortfall in gates 1 and 4 | Compare the same full graph and inputs with the reference; check spike IDs/ticks, state tolerances, sleep/wake and chunk continuity; measure CPU, memory and coupled-clock performance |
| Learning from intake | Experimental extension once repeated sensory responses work, supporting gates 1 and 3 | Cue-specific changes that persist; learning-off and unpaired-feedback controls; improved behavior on unseen placements, without giving the controller source coordinates |
navis is an analysis tool in this plan. Use it on the small set of candidate neurons before considering larger datasets. NBLAST can suggest morphological matches; it does not establish steering function. Keep skeleton/mesh versions, coordinate transforms and units explicit. Aligning two brain spaces does not make their neuron IDs interchangeable, and the default neuPrint hemibrain example is not our FlyWire 630 specimen.
The C/WASM candidate addresses computation. The useful idea is to skip integration when a bound proves a neuron cannot reach threshold before another input, then advance its state when needed. Preserve every connection and the coupled clock. Faster execution does not resolve persistent activity or missing direction; verify resource bounds and numerical agreement before replacing any solver.
Learning is a separate model experiment. The proposed feedback is the measured replenishment of the reserve currently in deficit, associated with recent sensory activity. This is an engineering hypothesis. Specify the plastic synapses, feedback rule and limits before testing, preserve the fixed reference assays, and require the learned response to work with continuous state and the full graph. An isolated circuit assay may diagnose learning but cannot close the foraging gate.
Use a dedicated fixture with one live fly, initial 70% reserves, reactive senses on and no source touching the mouth. Use Place in world, never Offer at mouth, for seeking tests. Save reachable left/front/right positions 1 cm from the initial body on an unobstructed surface within the existing placement limit.
The initial release targets are intake within 60 simulated seconds per source and at least 0.5× physical/wall-clock speed over the active sequence. These are targets, not measured performance. Test each position with three fixed seeds; save all nine outcomes and require all to pass. Fix coordinates, seeds and thresholds before tuning; document any necessary revision.
| Case | Pass condition |
|---|---|
| Apple at a distance | Sensor changes → neural commands → physical approach → feeding; energy transfers from the finite portion until it empties |
| Water at a distance | The same chain transfers water; that intake creates no energy |
| Apple → water → new apple | One life completes all three, with physical flight and landing during the sequence; no body or neural-state reset |
| Move/remove the source | Subsequent readings and commands respond; no pursuit of cached target coordinates |
| Disable relevant senses | Source-directed behavior is lost in the comparison; no intake while senses are disabled |
| Block neural motor commands | Sensory activity remains measurable, commanded locomotion disappears and distant food is not consumed |
| No food / no water | Each deprivation independently causes terminal death; no further neural actuation or automatic revival |
| Pause / Stop / New life | Freeze coupled time and reserves / cancel controller authority without automatic retry / restore reserves while preserving physical pose and objects |
Contact, baseline, bitter and conservation checks remain component regressions. They cannot substitute for this suite. Record body position, local sensors, neural readouts, actuator commands/forces, contact and resource transfers on one timeline, bound to code/data/controller hashes. Include a continuous capture of flight, landing, intake and departure; a screenshot is insufficient.
Measure physical/wall time, CPU, peak RAM and physics warnings for the combined workload. Keep one animal, 2 CPUs / 1 GiB and the existing 0.9-core physics and separate 0.15-core antennal duty budgets. Report actual rendering FPS separately from physics speed; target 30 FPS while moving, with no idle/hidden-tab draws. Check desktop and 390×844 controls. A guard trip is a failed run.
navis — inspected commit cb9a591. It provides morphology analysis, NBLAST, template transforms and visualization, with Rust-accelerated functions. Its coordinate guidance warns that comparisons across incompatible brain spaces can return plausible but incorrect results. The neuPrint interface requires a dataset selection and authenticated access. Plan offline analysis in Docker with selected dependencies; keep Three.js for the browser and flybody for physical motor execution. Blender exports can wait for presentation work.
hae — inspected commit edb1532. Its C core contains event-driven integration and an added dopamine-gated plasticity rule. The roughly 19 KB WASM file is executable code; graph data and learned weights are loaded separately. The conditioning experiment is a useful reference for paired and control odors, not a reproduced result here.
The character reader uses v783, disables outgoing transmission outside the mushroom-body circuit and resets neural state for each image while retaining learned gains. It assigns output groups to letters and receives the correct label during error feedback. This does not demonstrate a continuously operating full brain finding resources. Our separate v630 odor measurements also show persistent activity; the releases and protocols remain distinct.
The main scene uses predefined writing strokes with inverse kinematics, programmed food approach and interpolated flight. Its worker uses writing speed to drive displayed motor-neuron activity; that display does not prove those neurons caused the movement. These animations do not replace our neural-command-to-physical-actuator checks.
TMNF-C — inspected commit eb6be04. Its pathway audit is useful for checking bilateral anatomical support before interpreting directional readouts. Its separate mushroom-body learner and visual reflex driver provide experimental methods, with substantial differences from our full-graph FlyWire 630 model. The learner receives vehicle/route state; the visual driver uses engineered readouts, and replay modes can show neural activity while recorded commands control motion. See the source review, reported results and reuse boundaries. Apply the audit method to gate 1 first; learning and vision retain their existing place in the plan.
Any reuse requires pinned code/data, dependency and license review, bounded isolated execution and the relevant numerical and behavioral comparisons.
Touch/escape, learned vision, a reconstructed VNC, multiple flies, city expansion and texture polish wait until this loop passes. Earlier city ideas are a historical backlog. Additional diagrams, recordings or interface polish do not close a failed behavior gate.