Skip to content

Repository files navigation

CHIMERA

CHIMERA

A digital lifeform powered by the real Drosophila larva connectome

Built from 1,373 neurons and 22,400 synapses of an actual fruit fly larva brain
(Winding et al., Science 2023), connected to a MuJoCo physics body.

CHIMERA neurons synapses


What it does

  • Type anything → Qwen 0.5B parses natural language → sensory channels activated
  • 1,373 LIF neurons simulate the real connectome → motor signals emerge
  • MuJoCo body responds: walking, curling, head scanning, trembling
  • Response language matches input language (Korean / English / Japanese / Chinese)
"danger"   →  pain circuit fires       →  retreat
"위험해"   →  nociception circuit fires →  body curls, abdomen raises
"먹이냄새" →  chemical circuit fires    →  head scans left-right
"fly"      →  touch_front activates    →  forward locomotion

Language support: English (default), Korean (한국어), Japanese (日本語), Chinese (中文).
The response language automatically matches your input language.


Install

Requirements

  • Python 3.10+
  • Windows / macOS
curl -O https://raw.githubusercontent.com/caparison1234/chimera/main/setup.py
python3 setup.py
python3 chimera_app.py

setup.py automatically downloads:

  • Source files from GitHub
  • Connectome data (Winding et al., Science 2023)
  • Qwen2.5 0.5B GGUF from HuggingFace (~400MB)

Manual download (if setup.py fails)

Connectome:

  1. https://github.com/brain-networks/larval-drosophila-connectome
  2. Download ZIP, extract
  3. Copy data/Aw.csvchimera/connectome/real_adjacency.csv
  4. Copy data/nodeMeta.csvchimera/connectome/real_neurons.csv
  5. Run python chimera_load_connectome.py

Qwen model:

  1. https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF
  2. Download qwen2.5-0.5b-instruct-q4_k_m.gguf
  3. Place at chimera/models/Qwen2.5-0.5B-Instruct.Q4_K_M.gguf

Without the LLM, CHIMERA runs in rule-based mode and is still fully functional.


Build .exe (Windows)

build.bat

Produces dist/CHIMERA/CHIMERA.exe — distribute the entire dist/CHIMERA/ folder.


Architecture

Text Input
    ↓
Qwen 0.5B  (natural language → sensory JSON)
    ↓
LIF Neuron Simulation  (1,373 neurons, Winding 2023 connectome)
    ↓
Motor Signals  (forward / backward / curl / eat / tremble / wing)
    ↓
MuJoCo Physics Body  (12 actuators)
    ↓
First-person experience sentence  (neuron-type based translation)

Brain data

Winding, M. et al. (2023).
The connectome of an insect brain.
Science, 379(6636). https://doi.org/10.1126/science.add9330


Project roadmap

Phase Status
GENESIS — PPO locomotion Done
CHIMERA — Real connectome brain Done
AION Alpha — Reaction-diffusion layer Planned
AION Beta — Evolutionary body Planned
AION Release — Hebbian learning Planned

About

Drosophila larva connectome (1,373 neurons, 22,400 synapses) drives a real-time MuJoCo physics body. Sensory input activates biologically accurate LIF neural circuits. Based on Winding et al., Science 2023. Built with Python, MuJoCo, and Qwen2.5 LLM.

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages