Construction & limitations

01 / CONSTRUCTION

What the replay does

The game renders a forward flight course. Each frame is reduced to a small 48 × 27 image that stands in for a fly-facing visual sensor. Independent horizontal and vertical categorical policies choose movement from retinal ring error, confidence, apparent size, motion, and ring phase. A separate binary policy maps retinal target alignment to a front-limb firing event. Ring progress and target hits train their respective policies without sharing rewards. Stationary mode repeats one course for learning tests; randomized mode varies the ring layout to test transfer.

The 3D body is a browser display based on simplified NeuroMechFly body meshes. The small brain panel now uses the low-poly JRC2018U adult Drosophila atlas surface, with grouped visual, motor, front-limb, and dopamine-like activity markers layered on top.

02 / LIMITS

What it does not claim

• It is not a full Drosophila connectome simulation.
• The four visual channels are aggregated signals, not individual ommatidia or reconstructed visual neurons.
• The JRC2018U surface is an atlas average; the activity locations are still schematic pathway overlays, not measured firing locations.
• The clipped policy-gradient learner, value baselines, elite memory, and dopamine-like reward display are machine-learning abstractions, not demonstrated biological dopamine plasticity.
• The firing policy is a machine-learning abstraction of a front-limb decision, not a reconstructed leg premotor circuit. A small target-alignment movement is enabled only in course gaps and is intentionally capped to limit interference with ring steering.
• Training episodes remain stochastic. The moving average should trend upward, but individual episode scores are not expected to increase monotonically.
• The body animation has no aerodynamic, muscle, contact, or wingbeat physics.
• The body and brain references combine assets from different fly datasets and should not be interpreted as one exact individual.

03 / NEXT LAYER

What would make it more authentic

A stronger experiment would map retinal samples onto selected visual neurons, propagate activity through a curated visual-to-descending pathway, and drive the game from recorded or simulated motor outputs. A browser-friendly version should use a carefully selected subset of neurons and cache only the geometry needed for display.

The full connectome is better suited to an offline or server-side simulation, with the website receiving compact activity traces and replaying them alongside the retinal image.