nftechie/doomfly

What it solves

DOOMFLY attempts to test whether a biologically accurate simulation of a fruit fly's brain (the MaleCNS v1.0 connectome) can be trained to survive in a virtual environment (ViZDoom), using a dopamine-gated plasticity rule to simulate learning from negative reinforcement.

How it works

The project integrates a biological connectome with a game engine. It works through a following loop:

1. Sensory Input: Game frames are converted into brightness and color inputs that stimulate specific modeled sensory neurons.
2. Neural Processing: Activity propagates through a graph of 166,700 neurons and over 25 million directed connections based on a biological reconstruction.
3. Motor Output: Specific neural activity is mapped to game controls (turning, moving, and firing) via a fixed interface.
4. Learning Mechanism: When the agent takes nonfatal damage, an artificial aversive input is sent to dopamine cells, which triggers a plasticity rule that modifies a small set of existing connections (KC$→$MBON11).
5. Persistence: Neural states and memory persist across rounds, while death triggers a new arena round.

Who it’s for

Researchers and developers interested in embodied intelligence, biological neural networks, and the intersection of neuroscience and AI agents.

Highlights

• Biological Fidelity: Uses the MaleCNS v1.0 connectome with over 25 million connections.
• Live Telemetry: Includes a monochrome spectator website for monitoring the simulation in real-time.
• Hybrid Approach: Combines a biological wiring diagram with engineered controller assignments and artificial reinforcement.
• Transparent Results: Explicitly documents negative results and failed validation gates for current training candidates.