Back to browse
Evolved cells navigate a maze with no training or fitness function

Evolved cells navigate a maze with no training or fitness function

by heavymemory·Mar 20, 2026·2 points·0 comments

AI Analysis

●●●BangerWizardryRabbit HoleZero to One

Cells navigate mazes with no fitness function — behavior emerges from simple rules.

Strengths
  • ~2000 lines of single-threaded C — constraint-driven development on a laptop.
  • No neural network, no backprop, no training — pure evolutionary emergence from mutations.
  • Gene network reads local inputs and writes registers — physical consequences follow.
Weaknesses
  • Video demo only — no interactive version to experiment with parameters yourself.
  • More demonstration than tool — fascinating but limited practical application.
Category
Target Audience

Programmers interested in emergent behavior and evolutionary simulations

Post Description

Single file C simulation. Cells on a grid eat soil, fight neighbours, reproduce with mutations. No neural network. No backpropagation. No fitness function. No pathfinding. Evolution runs and behaviour emerges.

Left panel is the ecology where evolution happens. Right panel is a maze. I pick an evolved organism and drop one cell into the maze. Some genomes fail. Some explore the whole thing. Zero control after injection.

The cells don't have functions like 'move left' or 'eat food'. Each cell runs a small evolved gene network that reads local inputs and writes to registers. Physical consequences follow from the register values. The cell doesn't know it's navigating. Its internal chemistry just happens to produce movement.

~2000 lines of C. Single thread. Runs on a laptop

Similar Projects

Gaming●●Solid

I turned the Pong Wars simulation into a multiplayer game

Turns the viral Pong Wars sim into an actual competitive multiplayer game with powerups.

CozyCrowd Pleaser
mayerwin
201mo ago