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

Evolving Walkers

Build a creature from nodes and bones, mark some bones as muscles, and let evolution discover a gait. Nobody tells it how to walk — walking is whatever travels furthest.
What is happening here?

Your creature is a stick figure of point masses (nodes) joined by rigid bones. Some bones are muscles: they rhythmically lengthen and shorten. Each muscle has three evolved numbers — how far it stretches, how fast, and its phase offset — together a little central pattern generator, the same kind of oscillator circuit that drives stepping in real spinal cords. Those numbers are the creature's genome.

A population of 30 genomes is simulated; fitness is simply how far the body's centre travels in a fixed time. The best are selected, recombined and mutated, and the cycle repeats — the champion of each generation is replayed live above. Early creatures flop and twitch; useful rhythms are stumbled upon and kept. What emerges is not designed — it is the gait that happened to win, which is why evolved walkers so often move in strange, unintuitive ways. This is locomotion as an optimization problem, the bridge from your evolutionary-robotics session to embodied learning.

Try: the same body often evolves completely different gaits on different runs (evolution finds many local optima); a body with one muscle can only hop, while three or four can develop a stride; too few nodes touching the ground and it just falls over — morphology constrains what behaviour is reachable.

The champion's controller — a central pattern generator

A pacemaker hub (left) drives one oscillator per muscle. Each node's dial shows its phase; the wire to it lights up as that muscle contracts. The orange muscles on the creature above pulse in time with these — the whole "network" is just these few rhythms, the way a spinal CPG drives stepping.

1 — Build the body

presets:
bone muscle ● node

2 — Evolve

Best distance per generation