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
Productivité · SKILL.md

simulation-environment

simulation-environment

User-facing Go1 Isaac Gym environment and configuration guidance: Cfg and config_go1 setup, terrain, observations and history, reward scales, domain randomization, assets, installation and backend diagnosis, and safe static validation.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: simulation-environment description: "User-facing Go1 Isaac Gym environment and configuration guidance: Cfg and config_go1 setup, terrain, observations and history, reward scales, domain randomization, assets, installation and backend diagnosis, and safe static validation." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Simulation Environment

Use this sub-skill when the user needs static/API/configuration help for the Go1 Isaac Gym environment: choosing and validating Cfg fields, applying config_go1, understanding terrain and observation construction, checking reward and domain-randomization surfaces, locating assets, or diagnosing installation and backend prerequisites.

Operating boundary

  • This is a static, API, configuration, and prerequisite-diagnosis skill. Isaac Gym Preview 4 is absent in the construction environment, so do not launch simulation, training, playback, rendering, or native environment construction here.
  • A successful PyTorch import or torch.cuda.is_available() result is not evidence that Isaac Gym or the simulator runtime works.
  • Treat source evidence as already distilled into the bundled references. Do not tell future agents to open original repository files.
  • Route PPO/training/checkpoints to training-and-policy, actuator fitting or actuator-network behavior to actuator-network, and physical Go1 operation or hardware safety to robot-deployment.

Procedure

  1. Read configuration.md for the effective Go1 defaults, field names, configuration order, dimensions, and known static mismatches before suggesting edits.
  2. Read api-reference.md before describing constructors, wrapper behavior, or return values. Every API claim there is marked Isaac Gym-required and must not be presented as host-verified.
  3. Read runtime-prerequisites.md for installation facts, the separation between PyTorch and Isaac Gym, and the limits of this host. Use check_runtime.py for a read-only diagnostic when a repository root is available.
  4. Use validate_config.py to validate a JSON configuration summary or the bundled documented Go1 static defaults. It never imports Isaac Gym and is suitable for CI or a machine without the simulator.
  5. Use troubleshooting.md to classify a failure before changing configuration. Keep missing optional dependencies and unverified backend behavior explicit.

Bundled files

Handoff limit

The generated scope is static/API/configuration and prerequisite diagnosis pending Isaac Gym. It does not certify simulator execution, CUDA PhysX behavior, training quality, playback, checkpoints, actuator fitting, or robot hardware deployment.

Skills enregistrées