Skip to content
humanoidrobots.training

Tier 2 · NVIDIA GPUs

Train a policy

Training runs in mjlab on MuJoCo Warp, thousands of environments on one GPU. It needs CUDA, so it cannot run in this tab. This page writes the job; you run it, or send it to us.

Robots, train open
48
With stock mjlab task
1
Task template

Velocity-tracking reward on a flat plane. Stock env samples forward, lateral and yaw commands and randomises friction and pushes.

Run it yourself · mjlab

# NVIDIA GPU + CUDA required (mjlab runs on MuJoCo Warp). Does not run in a browser.
git clone https://github.com/mujocolab/mjlab.git && cd mjlab
uv run train Mjlab-Velocity-Flat-Unitree-G1 \
  --env.scene.num-envs 4096 \
  --agent.max-iterations 3000

# watch a checkpoint afterwards
uv run play Mjlab-Velocity-Flat-Unitree-G1 --checkpoint-file logs/rsl_rl/<run>/model_<iter>.pt
mujocolab/mjlabrun it on rented GPUs (SkyPilot)

Or ask us to run it · cloud GPU

There are no GPU workers attached to this site yet. This sends the job spec above to Hyperspawn as a request; a person reads it and replies. Nothing starts training when you press the button, and this page will never show a fake progress bar.

POST hyperspawn.io/api/v1/requests
Request body
{
  "kind": "training-run",
  "robot": "unitree/g1",
  "task": "Walk",
  "source": "humanoidrobots.training/train",
  "mjlabTask": "Mjlab-Velocity-Flat-Unitree-G1",
  "config": {
    "numEnvs": 4096,
    "maxIterations": 3000,
    "gpus": 1
  },
  "model": {
    "repo": "https://github.com/google-deepmind/mujoco_menagerie",
    "ref": "c96a32d28fb5da84da38c1da4d749e7a13212855",
    "path": "unitree_g1",
    "entry": "scene.xml"
  }
}