physicalai-bmi/forge-quadruped-gait
A learned quadruped walking gait, found entirely on-device in the browser by Evolution Strategies — no cloud, no autodiff, ~$0. This is the gait the Institute's Forge "Learn to walk" loop produces: a small CPG (central pattern generator) whose shape is a 10-parameter genome, optimized by a black-box ES against a headless MuJoCo-WASM forward-walk rollout.
- Baseline (hand-tuned default gait): 1.04 m in 3.2 s
- Learned (this genome): 1.42 m — +37%, and reproducible (deterministic physics)
- Method: antithetic ES, λ=16, 28 generations, in-browser MuJoCo-WASM
The genome (gait.json)
{ "gaitHz": 2.91, "stride": 0.183, "lift": 0.065, "standZ": -0.364, "x0": 0.002,
"duty": 0.485, "kpPitch": 0.235, "kdPitch": 0.040, "kpRoll": 0.419, "kdRoll": 0.013 }
gaitHz sets step frequency; stride/lift/standZ shape the elliptical foot path;
duty is the stance fraction; the kp*/kd* terms are a small closed-loop pitch/roll
balance controller the ES also tunes.
Run it
Load gait.json into the Forge quadruped and it walks — or watch the whole
learning loop discover it live at Forge → Learn to walk.
Reproduce this exact genome in ~a minute: the ES is deterministic given the seed.
Released CC-BY-4.0 by the Institute for Physical AI @ BMI. Companion to our released control policies (forge-arm-reach-bc, the nano-VLA family).