L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Déploiement et opérations · SKILL.md

humanoid-gym

humanoid-gym

Guide Humanoid-Gym users through PPO training, XBot-L environment customization, and MuJoCo sim-to-sim deployment.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: humanoid-gym description: "Guide Humanoid-Gym users through PPO training, XBot-L environment customization, and MuJoCo sim-to-sim deployment." disable-model-invocation: true metadata: disco-role: operating license: NO_LICENSE

Humanoid-Gym

Use this repo skill when a future agent needs to work with Humanoid-Gym: training or evaluating the bundled humanoid_ppo task, changing the XBot-L environment, or validating/exporting a policy for MuJoCo sim-to-sim.

Read this first

  • references/repo-provenance.md if you need to compare the current checkout against the snapshot used to build this skill.
  • references/installation-and-backends.md for Python, torch, Isaac Gym, MuJoCo, and CUDA readiness.
  • references/troubleshooting.md when setup, import, backend, or checkpoint selection fails.

Main routes

  • sub-skills/training-and-evaluation/ — PPO training, evaluation, checkpoint loading, and TorchScript export.
  • sub-skills/environment-customization/ — XBot-L config, rewards, terrain, and new robot/environment setup.
  • sub-skills/sim2sim-deployment/ — MuJoCo deployment of exported policies and asset/policy validation.

Verified package facts

  • Distribution name: humanoid
  • Version: 1.0.0
  • Registered public task in this checkout: humanoid_ppo
  • XBot-L policy dimensions: 705 observations, 219 privileged observations, 12 actions
  • The public play.py export path writes logs/<experiment_name>/exported/policies/policy_1.pt

Minimal install and check

Use a Python 3.8 environment for the repo's documented backend stack.

  1. Install this checkout in editable mode from a local clone.
  2. Install the runtime stack required by the selected route.
  3. Confirm import humanoid and python -m pip check.
  4. If you need train/play execution, confirm Isaac Gym Preview 4 is installed and importable.

Suggested smoke check:

python -c "import humanoid; print(humanoid.LEGGED_GYM_ROOT_DIR)"

If Isaac Gym is missing, stay in static/command-building mode and use the bundled helpers instead of claiming native training or evaluation succeeded.

Helper script

  • scripts/inspect_humanoid_gym_install.py — run this when you want a safe root-level install/resource/backend summary without launching Isaac Gym.

Notes for future agents

  • humanoid/scripts/train.py and humanoid/scripts/play.py are the user-facing PPO entry points, but they require the Isaac Gym backend for real execution.
  • humanoid/scripts/play.py exports the actor only and hard-codes render/export side effects unless the source file is edited.
  • humanoid/scripts/sim2sim.py uses MuJoCo viewer rollout and is best validated with the bundled sim2sim sub-skill.
Skills enregistrées