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
Données et recherche · SKILL.md

motion-tools

motion-tools

Inspect, convert, validate, view, and plot MimicKit motion, dataset, viewer, DOF, conversion, and log artifacts.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: motion-tools description: "Inspect, convert, validate, view, and plot MimicKit motion, dataset, viewer, DOF, conversion, and log artifacts." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MimicKit Motion Tools

Use this sub-skill when a task is about MimicKit motion artifacts rather than policy training: motion pickle files, motion dataset manifests, GMR or SMPL/AMASS conversion, view_motion visualization, character DoF diagnostics, and training-log plots.

Start here

  1. Read references/motion-format.md before writing or validating a motion file or dataset YAML.
  2. Read references/conversion-and-visualization.md for bundled converter commands, viewer/DOF workflows, and log plotting.
  3. If a workflow fails, route to references/troubleshooting.md before changing data or configs.

Bundled helpers

All helpers are self-contained and accept ordinary file paths. When a target MimicKit checkout should be made importable, pass its root explicitly with --repo-root; do not rely on the current working directory being the checkout.

Route elsewhere

  • Full train/test runner flags, backend installation, device selection, and simulator choice belong to the sibling runner-and-backends sub-skill.
  • DeepMimic, AWR, LCP, AMP, ADD, ASE, and other policy-training recipes belong to algorithm sub-skills.
  • Score-Matching Motion Prior training belongs to the sibling smp sub-skill; this sub-skill only covers generic motion/data conversion and visualization.

Verification limits to preserve

The generated guidance is backed by source inspection plus lightweight checks: CPU/CUDA torch imports, parser help for the GMR and SMPL converters, source compile checks, and tiny converter fixtures. Simulator-native viewing and DOF-test workflows still require an external supported simulator backend plus downloaded motions/models/assets before they can be fully validated.

Skills enregistrées