GRADY.OS

/ walkers

Walkers learn to walk.

Every creature has a tiny neural network for a brain, wired at random. None of them know how to walk. The ones that stumble furthest get to have children. Give it a few minutes.

Generation 0 has never taken a step

120 random brains. One body. Ten seconds each.

Or first.

Creature
Speed

Progress

Generation
0
Best ever
—
Creatures tested
0
Simulated
0 s

best of each generation median

Leaderboard last generation

  1. Nobody has finished a run yet.

Dynasties

Every founder starts a family. Watch a few take over.

    Inside a brain —

    Left: what it senses (a clock, its speed and height, where each joint is and whether it touches the ground). Middle: 10 neurons. Right: one output per muscle. Lime contracts, coral stretches.

    Hall of fame

    1. Records get a flag on the track.

    How it’s built

    Physics. Joints are point masses joined by springs: stiff ones for bones, softer ones for muscles that can shorten or stretch by a quarter. Gravity, damping, and ground friction worked out as impulses, 240 steps a second.

    Brains. Each creature gets a small neural network: a clock, its speed, its height and every joint’s position in; ten tanh neurons; one output per muscle. About 200 numbers decide how it moves.

    Evolution. No training data and no gradients: a genetic algorithm. The fastest three survive as they are, the rest of the next generation are children of tournament winners, with random mutations and the odd crossover.

    Speed. Turbo scores whole generations in parallel on web workers while the screen replays the champion. The physics is deterministic, so the run you watch is the exact run that was scored.

    1,008 lines of hand-written JavaScript and CSS, no libraries, no network requests. 12 tests passed when this page was built. Nothing you make leaves your browser.