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

model-editing

model-editing

Guide safe Menagerie MJCF editing, attachment composition, actuator gains, mirrored hands, bi-arm specs, and MJX-aware variants.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: model-editing description: "Guide safe Menagerie MJCF editing, attachment composition, actuator gains, mirrored hands, bi-arm specs, and MJX-aware variants." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Model Editing

Use this sub-skill when the task is about changing or generating Menagerie-style MJCF/XML content: attachment sites, arm/hand composition, actuator/keyframe variants, PD gains, mirrored hands, bi-arm assemblies, or MJX-specific model forks.

Route boundaries

  • Stay here for edit plans, local XML transformation patterns, actuator/keyframe updates, and the bundled read-only gain helper.
  • Route model choice, model-directory anatomy, scene-vs-model XML selection, and MJX availability questions to model-catalog.
  • Route direct mujoco loading, viewer use, compile/step debugging, and asset-path smoke tests to model-loading.
  • Route final formatting, changelog/license updates, gallery changes, and CI/check orchestration to contribution-maintenance.
  • Do not require future agents to run Menagerie source helper scripts from a checkout. Use the distilled workflows and bundled script in this sub-skill instead.

Inputs to collect before editing

  1. Base XML path(s), whether they are canonical model XMLs, scene XMLs, or variants.
  2. Requested output path and overwrite policy; default to a new file or copy, not in-place source mutation.
  3. Composition intent: attach end-effector, duplicate an arm, mirror a hand, add/remove a controller, or create an MJX fork.
  4. Keyframe policy: preserve all keyframes, expand only home, create a new keyframe, or remove stale keyframes.
  5. Validation budget and handoff target for loading/CI checks.

Read first

  • Model editing workflows for attachment sites, MjSpec composition, mirrored/bi-arm generation patterns, actuator variants, PD gain formulas, and the editing checklist.
  • MJX and variants for _mjx.xml, scene_mjx.xml, no-hand/with-hand/controller variants, and common MJX caveats.
  • Troubleshooting for missing attachment sites, keyframe length mismatches, invalid joint references, mirrored mesh winding, gain-helper failures, and MJX incompatibilities.

Bundled helper

Use scripts/compute_pd_gains.py when the task asks for Menagerie-style position actuator kp/kv estimates from effective joint inertia:

python scripts/compute_pd_gains.py /path/to/model.xml --saturation-angle-deg 10 --damping-ratio 1.0
python scripts/compute_pd_gains.py /path/to/model.xml --keyframe home --frequency-hz 2.0 --damping-ratio 0.9

The helper is read-only. It accepts arbitrary MJCF XML paths, can use actuator or joint force limits, and prints candidate gains without modifying the XML.

Validation signals before handoff

  • Every generated or edited XML is written to an explicit output path and the original source XML remains available for comparison.
  • Attachment and MjSpec composition plans include keyframe qpos/ctrl expansion or a decision to remove/regenerate stale keyframes.
  • Mirroring plans cover mesh vertex reflection, face-winding reversal, pose/default mirroring, identifier renaming, and at least one geometric sanity check.
  • MJX edits are kept in separate variant XMLs and do not silently replace canonical Menagerie models.
  • After the edit plan or local modification is complete, hand off compile/smoke testing to model-loading and final contribution checks to contribution-maintenance.
Skills enregistrées