They Copied a Fly's Brain Into a Computer.

Papers

• Complete adult fruit fly brain connectome — Nature (Oct 2, 2024):
  “A complete wiring diagram of the fruit‑fly brain” (special issue overview + package of 9 papers).
  Read: Nature news/overview, NIH Research Matters summary, FlyBase reference bundle
• Whole‑body physics simulation of a fruit fly (the body/platform) — Nature (Apr 23, 2025):
  “Whole‑body physics simulation of fruit fly locomotion” (MuJoCo-based, DeepMind + Janelia/Turaga Lab).
  PDF: Nature article • Code: TuragaLab/flybody on GitHub
• Connectome‑constrained controller that uses the real fly brain graph — arXiv preprint (Mar 8, 2026):
  “Whole‑Brain Connectomic Graph Model Enables Whole‑Body Locomotion Control in Fruit Fly (FlyGM).”
  Read: arXiv (HTML)

Why it Matters

They Copied a Fly’s Brain — And It Walked (No Training Required)

For decades, mainstream AI has assumed that learning creates intelligence: define a loss, run optimization, and behavior emerges. Biology often flips that script: build the right wiring first, then behavior follows. A series of studies around the humble fruit fly (Drosophila melanogaster) is pushing that idea from metaphor to working systems.

 

Step 1: Map the whole brain.
In October 2024, a global consortium released the first complete connectome of an adult fruit fly brain—about 140k neurons and >50 million synapses—as a publicly available resource. This Nature special issue (plus nine companion papers) provides the first at-scale wiring diagram that spans the entire adult fly brain, enabling researchers to reason from actual circuitry rather than coarse abstractions. Overview in Nature • NIH summary • Reference bundle

 

Step 2: Build a body the brain can drive.
To study real behavior, you need more than a brain—you need physics. In 2025, a team from DeepMind and Janelia/Turaga Lab published a high‑fidelity, physics‑based fly body in Nature. This platform reproduces realistic walking and flight in a MuJoCo simulator and is designed as an open, general framework for embodied neuroscience and AI. It’s the standardized “body” many groups now plug their neural models into. Paper (Nature, 2025) • Code (GitHub)

 

Step 3: Let the wiring be the policy.
In March 2026, researchers introduced FlyGM, an arXiv preprint that takes the exact brain connectome and represents it as a directed message‑passing graph—a neural controller whose static structure is identical to the real fly’s wiring. Hooked up to the physics‑accurate body, the system achieves stable control across diverse locomotion tasks without hand‑crafted architectures; critically, its structure outperforms random graphs or standard MLPs in sample efficiency and performance. The promise: start from biologically grounded topology, then (if needed) learn on top of it. FlyGM (arXiv, 2026)

 

Why this is a big deal

• Structure-first intelligence: These results suggest that richly structured, evolution‑shaped wiring can embed powerful behavioral priors—reducing or even eliminating the need for heavy training loops in certain regimes. Nature connectome overview • FlyGM
• Embodiment matters: The Nature body model anchors “brains” in believable physics, closing the loop between sensory input → neural dynamics → motor output → new sensory input. This enables rigorous, end‑to‑end tests of how neural circuits actually cause behavior. Nature body model
• Interpretability by construction: When your controller is the biological graph, every edge and node has a biological meaning. Causal analyses map naturally onto real neuron types and pathways—an attractive alternative to black‑box policies. Connectome issue • FlyGM