Sim Env: Open-Ended Evolution

What it is

Sim Env is a 2D world in which creatures with growing bodies and recurrent, plastic neural brains evolve with no goal other than staying alive and reproducing. There is no fitness function and nothing is scripted after birth: whatever gathers energy and leaves offspring spreads.

Light is the only energy that enters the world. Producers live on it, grazers eat producers, and predators hunt grazers. Seasons move the sunny band across the map, clouds drift, lakes and moisture decide where plants can grow, and random climate events (drought, ice age, heat wave, volcanic winter, plague, bounty) reshuffle who survives. The point is to watch macro evolution happen: species rising, taking over, crashing and adapting after a catastrophe.

This is a rewrite of an earlier browser version, Sim Env v1.

What evolves

Bodies	A growth program on an 11×11 grid. Creatures are born as a single core and grow (and regrow) cells as energy allows. 14 cell types — tissue, armour, fin, eye, mouth, spike, photosynthetic, fat, toxin gland, shield, jet, light sensor, gonad — and placement matters: eyes see where they face, spikes hurt what they touch, fat insulates.
Brains	144 inputs → 32 leaky recurrent units → 14 outputs. Time constants and neuromodulated Hebbian plasticity are part of the genome, so how much a creature learns during its life evolves too.
Senses	8 vision sectors, smell, pheromones, temperature, hydration, a compass and a clock, plus memory of where food and threats were last seen.
Physiology	Warm- or cold-blooded, preferred temperature, fat, a gut for gorging, torpor. Frostbite, heatstroke, thirst and disease are real causes of death.

The engine

Background “training worlds” (islands, a duel league, a species school) run on the same engine and send adapted genomes back into the main world. A creature bank keeps the best organisms of every run and can found new worlds.

Watching it live

The desktop app is for experiments. For an always-on world, the engine runs headless in evo_serve on a small machine on the LAN, and browsers do all the drawing, like clients of an online game.

Part	Role
Sim host (LAN)	Runs the world, saves on every stop or update, and sends the world state as compact binary frames over a tiny built-in HTTP server
nginx (web server)	Public, read-only reverse proxy with a shared cache; LAN-only controls are blocked
Browser	Decodes the frames, interpolates motion, and renders the map, population masses, cell bodies, charts and an inspector on a canvas

Measurements

Measured on the live setup (medium world, 4 frames per second) and on a local nginx 1.24 test bench:

	Result
Bandwidth per viewer, zoomed out	0.24–0.3 Mbit/s (7–9 KiB per frame)
Bandwidth per viewer, zoomed in	0.4–1.0 Mbit/s
100 simulated viewers through nginx	all at 4.1 frames/s, 0 errors; 8,421 of 8,521 frame requests were cache hits, so the host saw the same load as for one viewer
Frame size, 5,178 creatures	88 KB raw → ~20 KB as a gzipped delta
Decoding	every delta frame compared against a fresh keyframe: identical
Update	build, 40 tests and restart on the host in ~35 s; the world saves and resumes, open pages reload themselves

The first host is a spare laptop. Running uncapped on four cores it sat at 92 °C, so it now runs at a fixed 60 ticks per second (58–66 °C). The same cap will matter on the Orange Pi.

Status

The engine, the desktop app and the live pipeline work, and the public page runs from a LAN host behind nginx. Still to do: move the host to an Orange Pi 4 Pro and measure it there (the brain step has no ARM NEON path yet), and bring more of the desktop inspector, such as the live brain, to the browser.