Skip to experiment
SYN·CO·RE LAB AI Experiment · PSY3230

Evolutionary Robotics

Nobody programs these robots. A population of random nervous systems is bred — the ones that find food reproduce, with mutation. Behavior that looks like appetite, pursuit, even foraging strategy emerges from selection alone.
What is happening here?

Each robot is a Braitenberg vehicle: two light sensors (whiskers), two wheels, and between them a tiny neural controller — six numbers deciding how strongly each sensor excites or inhibits each wheel. Those six numbers are its genome. Generation zero is pure noise: robots spin, tremble, drive into walls. Every 20 simulated seconds, the robots that ate the most food are selected, their genomes recombined and mutated, and a new generation starts. Watch coordinated food-seeking appear within a dozen generations — typically a crossed excitatory wiring, exactly the circuit Braitenberg drew for the vehicle he named "aggression". You can inspect the champion's brain on the right while it evolves.

Valentino Braitenberg called this synthetic psychology: his point was that observers cannot resist describing such trivial circuits as wanting, fearing, loving — psychology's attribution problem in its purest form. Evolutionary robotics (Floreano, Nolfi) adds the second lesson: design by selection rather than by intention. The "design" lives in the fitness function, and the solutions are often not what an engineer would have written — they merely work.

Experiments: raise the mutation rate to 100% and watch evolution dissolve into drift; set it to zero after a few generations and watch the population converge and stagnate. Selection needs variation — but not too much.

Champion's brain (best robot, highlighted in the arena)

Fitness over generations (green: best, grey: average)