FlyGenesis: a fly building bodies to solve physical worlds

After putting MaleCNS inside a 3D navigation loop, the next question became more interesting: what if the fly did not receive a fixed body?

What if it participated in a process that creates bodies, tests them in physics, preserves the useful ones, and progressively makes the world harder?

The first demonstrator is embedded below. It reuses the same FlyDoom assets: the compact 165,122-neuron MaleCNS connectome and the anatomical map of sensory and descending neuron groups. The site does not ship a second copy of the brain.

The loop#

physical world
   ↓
synthetic sensors
   ↓
MaleCNS visual + olfactory neurons
   ↓
165,122-neuron recurrent connectome
   ↓
descending neurons
   ↓
motor adapter
   ↓
mutable body
   ↓
distance, energy and success
   ↓
selection + mutation

The body starts simple. Mutations can change limb length, angle, phase, stiffness, adhesion and actuator gain; some mutations add or remove limbs. A body that solves the current world better can enter the lineage.

When the lineage reaches the goal, the physical curriculum advances: flat ground, a step, a valley, and mixed terrain.

The “proteins” are still a functional alphabet#

The current demo deliberately separates morphology search from molecular simulation.

Actomyosin, collagen, elastin and adhesion are functional abstractions for actuation, structure, compliance and grip. We are not yet simulating amino-acid sequences, folding, molecular dynamics or real protein fabrication.

That is intentional. First we want to see whether there is an interesting search problem at the level of morphology + control + physics. If useful body plans emerge, a later layer can ask which real proteins or materials could implement each function.

What actually comes from MaleCNS#

The demo directly loads:

• /flydoom/malecns_l3_compact.mcns
• /flydoom/malecns_circuit.json

The browser kernel uses the same compact recurrent representation as FlyDoom. Body and terrain signals are projected into mapped sensory populations, while descending-neuron activity is read by the motor adapter.

So the recurrent topology is MaleCNS. The sensors, body model, adapter, fitness function and evolutionary algorithm are synthetic. This is an architecture demonstrator, not evidence that a biological fly can design robots.

The experiment behind the toy#

The long-term goal is not to reward larger bodies. Complexity should survive only when it buys physical capability.

The scientific comparison comes next: run MaleCNS, degree-preserving randomized reservoirs and small learned networks through the same morphological search, with matched budgets, seeds and curricula.

The question then becomes:

Does biological connectome dynamics provide any measurable advantage when brain, body and physical curriculum co-evolve?