[siddhss5]

Summary

Adds one convention for MuJoCo geom and site groups across the workspace's asset models, plus a checker and an integration test that enforce it. Today groups are assigned ad hoc: robots follow MuJoCo Menagerie, but the ABH hands, every prl_assets object, and the sites of different robots disagree. So viewers show different kinds of content by default depending on which model loaded.

Changes

• docs/mujoco-groups.md, the canonical convention, follows Menagerie and MuJoCo's default visibility (groups 0–2 shown):
  • Geoms:
    • 0: renders and collides (floors, tables, simple props).
    • 2: visual meshes.
    • 3: collision geometry, including physics-disabled safety and planning volumes.
    • 1, 4, 5: unused.
    • Collision geometry is hidden with group 3, not alpha 0, so toggling group 3 reveals it.
  • Sites:
    • 0: task frames (grasp, end-effector, tool tip, feeding target).
    • 4: engineering frames (attachment and mount points, F/T and IMU, camera frames, force pads, region markers).
  • A debug preset enables specific groups, for example {0, 1, 2, 4}, not all of them.
• scripts/check_mujoco_groups.py MODEL.xml ... reports each geom and site that breaks the rules and exits non-zero. It reads what the convention can determine mechanically:
  • collision versus physics-disabled from contype/conaffinity;
  • "beside a visual mesh" per body;
  • resolved alpha, using MuJoCo's material rule;
  • site groups limited to {0, 4}.
• tests/integration/test_mujoco_groups.py runs the checker on every model in geodude_assets, ada_assets, and prl_assets. Models that do not comply yet are strict expected failures linked to their issues. A fixed model therefore makes its entry fail until the entry is removed, so the list cannot go stale.
• README links the convention.

Survey that motivated it

33 models were checked:

Content	Usual group	Exceptions
Robot visual meshes	2	ABH hands: 1
Robot collision geometry	3	none
prl_assets objects	none	everything in 0, collision proxies hidden with alpha 0
Sites	none	ADA all 0; Geodude/UR5e grasp 0, sensors and attachments 4; ABH 3; worktop 2

No code in mj_manipulator, geodude, ada_mj, ssrobot, or sscbirrt reads group numbers, so this changes only what viewers show.

Asset fixes (merged)

• Follow the workspace MuJoCo group convention prl_assets#19 (closes prl_assets#18): 16 objects. Visuals moved to 2, and collision proxies to 3 with a color instead of alpha 0.
• Follow the workspace MuJoCo group convention geodude_assets#29 (closes geodude_assets#28): ABH hand visuals 1 → 2, their force-sensor sites 3 → 4, and the worktop site 2 → 4.
• Put ADA's engineering frames in site group 4 ada_assets#36 (closes ada_assets#35): attachment, camera, and sensor frames moved to site group 4.

KNOWN_VIOLATIONS is now empty. It stays as the mechanism for any future non-compliant model.

Testing

Against the asset repos' current main, the integration test gives 26 passed (all models). Workspace Integration CI is red for reasons that predate this PR (#173). The new files pass ruff check and ruff format.

[siddhss5]

CI here is red for reasons that predate this PR: sscbirrt's tests are missing hypothesis, and the ada_mj and mj_viser gitlinks are stale. Both are tracked in #173, to be fixed after the asset group issues. The new test passes locally against the asset repos' main: 6 passed, 20 strict expected failures.