PERSONAL / CREATURE LAB
Creature Lab
One fruit fly. Three experimental worlds: a 3D terrain playground, a neural maze, and a Pac-Man arena that learns from rewards.
Personal side project · Simulation & learning
My contribution
Built a unified light/dark hub, 3D replay and terrain controls, neural telemetry, an arcade-inspired ghost engine, and a reward-trained motor readout with paired evaluation.
Try it · mini version
See what it can do
Personal side project · Fly training, physics & neural simulation
What this project does
Creature Lab brings three fruit-fly experiments into one responsive home: a 3D walking world, a FlyWire neural maze, and Fly Pac-Man. A shared green theme connects light and dark modes, fullscreen viewers, replay controls, and visible evidence.
Inside the project
- Explore the 3D terrain lab with orbit, middle-drag pan, zoom, editable ramps and recorded MuJoCo movement.
- Compare recorded neural-maze behavior with intact wiring, disconnected edges, silenced descending output and untrained controls.
- Train an artificial Pac-Man motor readout through rewards. After 20,000 additional training games, the validation-selected checkpoint improved from 60/200 to 97/200 wins on the same fresh test games: 30% to 48.5%.
- Use the public showcase for recordings; run the Python gateway for live keyboard play, fresh ghost seeds and optional training through a configured tailnet server.
Skills & practical experience
Practice connecting scientific simulation to an accessible interface, implementing reproducible game logic, evaluating continued learning on paired tests, and documenting the boundary between measured anatomy and engineered control.
The gallery contains actual interface captures. The 19-second tour combines those captures with recorded body-physics footage. The 30% and 48.5% results compare two trained readouts on the same 200 seeds; only the readout learns, and these are simulated experiments rather than a living fly brain.
Project context
Measured FlyWire connectivity is used in computer simulations. Anatomical wiring stays fixed; an artificial motor readout learns. The public showcase plays saved experiments, while the local Python app supports live Pac-Man and optional tailnet training.
Project media




