Using a 222-Neuron Fruit Fly Connectome to Playtest My Arcade Shooter: What 471 Waves Revealed About Late-Game Balance
retained text
Scientists open sourced a full male fruit fly brain so I hooked it to my game and had it play. Got valuable insights. Just wanted to share cause I found it genuinely exciting and helpful. I recently made some improvements to my game so I wanted to stress test but playing for ages was not the best option (I do ofc self test) so I thought "scientists recently open sourced male fruit fly brain, how about I use that ?" and so I tried. It actually worked out very well and gave me insights which helped me fix and improve the overall game in many ways. [https://i.postimg.cc/CLFm4CK7/1789678885689.jpg](https://i.postimg.cc/CLFm4CK7/1789678885689.jpg) The setup connects Godot 4 with a Python-based biological neural simulator running in real time: **1. Engine Communication:** Godot's \`PacketPeerUDP\` sends normalized 3D world target coordinates and temperature metrics to Python each frame, and receives back a virtual aim delta \`(dx, dy)\` and trigger state. **2. Visual Processing (Lobula):** Screen-space target coordinates are converted into sensory inputs for the optic lobe's Lobula neurons. **3. Steering Compass (Central Complex):** Biological insects use the Central Complex as a navigation and orientation compass. Here, it computes the angular error between current aim and the target, outputting motor steering signals. **4. Decoy Filtering (Mushroom Body & MBON):** In late boss waves, the sun spawns decoys (\*Heat Mirages\*). The Mushroom Body output neurons filter visual noise to track only the primary heat source. **5. Dopamine Loop (PAM / PPL1):** In Drosophila biology, PAM neurons signal positive reward (water hits target), while PPL1 neurons encode aversive punishment (sun heat increases). Dopamine dynamically modulates synaptic weights during the run. **The results:** The agent started with roughly **36% aim accuracy** in the first 10,000 frames, constantly drifting and overshooting whenever the sun darted across the horizon. However, as dopamine-driven synaptic plasticity tuned the weights, accuracy crossed 70% by frame 25,000. By wave 400+, it reached a **99% instantaneous tracking lock-on**, with tracking error converging to dx: 0.00, dy: -0.002. The simulated fly survived all the way to **Wave 471 (38,650 frames / \~22 minutes of uninterrupted bullet-hell firing)**, racking up +28,267 cumulative dopamine without dying a single time. **What this revealed about the game:** **1. Uncapped Buff Compounding:** After dozens of boss wave drafts, stacking power and cooling multipliers created power creep. By Wave 200+, the sun was cooling down virtually the instant it spawned. **2. The "Hold the Trigger" Problem:** No matter how fast the sun moved or how much health it gained, keeping the crosshair locked and holding down primary fire neutralized all tension. The game devolved into a passive math-check instead of requiring tactical skill. I have applied several fixes and improved the game thanks to results I got from this experimentation. I always test my game and several users have already given me valuable insights, this was done more as experimentation. Thanks for reading the post\~
document fetched 2026-09-19 00:26 · canonical https://reddit.com/r/gamedev/comments/1wjtoir/using_a_222neuron_fruit_fly_connectome_to