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

locomotion-manipulation

locomotion-manipulation

Use high-level dm_control locomotion, soccer, mocap, walker, prop, and manipulation task families.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: locomotion-manipulation description: "Use high-level dm_control locomotion, soccer, mocap, walker, prop, and manipulation task families." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

dm_control locomotion and manipulation router

Use this sub-skill when the task is about selecting or using built-in high-level dm_control manipulation, locomotion, soccer, walker, mocap, arena, or prop families. These are Composer-backed environments and components; they can be stepped with the same dm_env loop style as suite tasks once an environment is constructed.

Route first

  • For ready-made manipulation registry tasks, read references/manipulation-reference.md to choose features versus vision, validate names/tags, load with dm_control.manipulation.load, and smoke-test specs.
  • For locomotion examples, walkers, arenas, soccer, mocap/reference-pose tracking, and asset/data caveats, read references/locomotion-reference.md before constructing or importing heavier task families.
  • For common failures around unknown names/tags, feature/vision confusion, missing HDF5/labmaze/assets, expensive demos, and rendering requirements, read references/troubleshooting.md.
  • Run scripts/list_locomotion_manipulation.py to list installed manipulation tasks/tags and verify that common locomotion example modules import; add --smoke-manipulation for a one-reset/one-step manipulation smoke.

Boundaries

  • Generic Control Suite tasks and benchmark loops belong in ../suite-rl-workflows/SKILL.md.
  • Building new Composer tasks/entities, custom observables, or reusable robot abstractions from scratch belongs in ../composer-environments/SKILL.md.
  • Rendering backend operations, viewer windows, and pixel-render troubleshooting belong in ../rendering-viewer-assets/SKILL.md.

Quick decision checklist

  1. If the user names a manipulation task or asks for Jaco arm/hand tasks, validate it against dm_control.manipulation.ALL and load it with dm_control.manipulation.load(name, seed=...).
  2. If the user wants low-dimensional observations, prefer manipulation names ending in _features; if they want camera observations, use _vision and verify rendering support separately.
  3. If the user names CMU humanoid, rodent, soccer, maze, corridor, bowl, target, or mocap/reference-pose tracking, treat it as locomotion/Composer-backed and inspect specs with a short reset/step before any long rollout.
  4. If the task requires new arenas, reward functions, entities, or observables beyond choosing/configuring built-ins, route to Composer rather than extending this catalog inline.
Skills enregistrées