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

rendering-viewer-assets

rendering-viewer-assets

Choose and troubleshoot dm_control rendering backends, viewers, pixel observations, camera APIs, and optional Blender exporter workflows.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: rendering-viewer-assets description: "Choose and troubleshoot dm_control rendering backends, viewers, pixel observations, camera APIs, and optional Blender exporter workflows." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

rendering-viewer-assets

Use this sub-skill when a task needs dm_control offscreen rendering, pixel observations, camera control, the interactive viewer, or safe guidance for the optional Blender MuJoCo exporter.

Do not use this sub-skill for non-rendering Control Suite rollouts; route those to ../suite-rl-workflows/SKILL.md. Route raw MJCF/MuJoCo model construction to ../mjcf-mujoco-models/SKILL.md, and custom Composer environment/task lifecycle work to ../composer-environments/SKILL.md.

First decisions

  1. Identify the rendering target:
    • Offscreen RGB/depth/segmentation arrays: use physics.render(...) and probe MUJOCO_GL before building the workflow.
    • RL pixel observations: validate a backend first, then wrap with dm_control.suite.wrappers.pixels.Wrapper.
    • Interactive window: use dm_control.viewer.launch(...); it requires GLFW and a real display.
    • Blender export: treat as optional, external-Blender, documentation-only unless the user explicitly accepts Blender add-on installation risks.
  2. Choose a backend before importing dm_control rendering modules. If MUJOCO_GL is unset, dm_control tries GLFW, then EGL, then OSMesa. Explicit values are glfw, egl, and osmesa; EGL also accepts MUJOCO_EGL_DEVICE_ID for selecting a GPU device.
  3. Do not promise that every backend is installed. During construction, a headless EGL render probe passed, OSMesa failed because the OSMesa/OpenGL library was missing or unusable, and GLFW failed in a display-less headless session.
  4. Install public packages only: use pip install dm_control for released packages, or pip install git+https://github.com/google-deepmind/dm_control.git for unreleased source snapshots. Editable installs are not supported by dm_control.

Bundled operating files

Minimal operating pattern

python scripts/render_backend_probe.py --backend egl --height 48 --width 64

If the probe succeeds, use the same backend selection for the downstream render/pixel-observation process. If it fails, follow the failure-specific route in references/troubleshooting.md instead of changing unrelated suite, MJCF, or Composer logic.

Skills enregistrées