Verify MuJoCo

Verify MuJoCo

Use this when forgecad export mjcf ... is part of the deliverable. A model is not sim-ready just because forgecad check simready passes or the MJCF file loads: it must be loaded in MuJoCo, stepped under gravity, driven with the intended controls, contact pairs inspected, and rendered from useful views.

Routing: geometry-only visual inspection -> forgecad-inspect-model; model authoring/API questions -> forgecad; building a new model -> forgecad-build-model.

Definition Of Done

1. Export from the exact source file.

   rm -rf /tmp/forgecad-mjcf && mkdir -p /tmp/forgecad-mjcf
   forgecad export mjcf path/to/model.forge.js --output /tmp/forgecad-mjcf
   

2. Load the generated scene in MuJoCo. Use scene.xml, not only the model XML, so the floor/camera/package context is included.
3. Check the root behavior. If the model has a free root, it needs real support/contact geometry or an explicit fixed-root export path. Do not hide a floor failure by turning the whole shell into a giant bounding box that blocks moving internals.
4. Check initial poses numerically. For mechanisms, compute the expected body/link axes from the design source and compare them to MuJoCo xmat/xpos. Joint ref/default and connector frames can disagree with visual intuition.
5. Keep meaningful collisions. Do not mark moving functional parts Sim.collider.none(...) just to make motion pass. If full visual mesh contact is unstable, use a physically defensible simplified collider or proxy and state what physical surface it represents.
6. Run the intended control with numeric acceptance criteria. Define the expected signed post-settle joint travel before running the test: e.g. "this drive should rotate the drum -0.04 to -0.06 cycles, then stop near zero velocity" or "this wheel should move at least +1 cycle and keep spinning". Use cycles for revolute/indexing mechanisms when that is easier to reason about, and radians for direct MuJoCo qpos checks. A controller that only jitters, moves the wrong direction, overshoots through a stop, or eventually jams after skipping the intended state is a failure even when the process exits 0.
7. Inspect contact pairs. Contact names should match the physical story: floor/support, card/card, wheel/ground, stop/follower, etc. Contacts against a filled-in AABB, hidden fixture, or unrelated side plate usually indicate bad collider selection.
8. Render evidence. Save initial, settled, and driven frames from views that actually show the moving parts. If the model orientation is not obvious, generate a labeled camera preview grid first, inspect it, then rerun with the azimuth/elevation that shows the functional face. Do not report GIFs/frames before visually confirming they are not from the back, underside, or an occluded side.

Helper Script

This skill ships a MuJoCo smoke verifier:

uv run --python 3.11 --with mujoco --with pillow \
  python <this-skill-dir>/scripts/mujoco_verify.py /tmp/forgecad-mjcf \
  --settle-seconds 2 \
  --seconds 8 \
  --actuator drum_velocity=-0.75 \
  --watch-joint drum_joint \
  --expect-drive-cycles drum_joint=-0.06:-0.03 \
  --expect-final-qvel drum_joint=-0.02:0.02 \
  --render-dir /tmp/forgecad-mjcf/verify \
  --camera-preview-grid

Use --actuator name=value more than once for multi-actuator models. Use --expect-drive-cycles joint=min:max to assert signed final-minus-settled revolute travel in cycles/turns. Use --expect-drive-delta joint=min:max when you want raw MuJoCo qpos units instead. Use --expect-final-qvel joint=min:max to assert terminal velocity when the mechanism should stop or continue at a bounded speed. The script prints JSON with root drift, initial-to-final joint delta, post-settle drive delta, derived cycle counts, final velocities, expectation ranges, and top contact pairs, then writes PNG frames if --render-dir is supplied.