fix(screwcap): regenerate bottle_cap + juice_shot USD — mount offset applied once, MJCF contact excludes carried over

#38
by Everloom - opened

Regenerates both screw-cap USDs. The published blobs put the cap in the wrong place on IsaacSim while MuJoCo, which never reads them, stayed correct — hand_track.open_bottle_cap RT2 frames showed a cap floating above its bottle.

Two independent defects, fixed in RoboVerse/scripts/mjcf_screwcap_to_usd.py (WUJI_SimBench PR #124):

  1. The mount offset was applied twice. cap hangs off an intermediate cap_screw link (PhysX articulations must be trees); that link's origin already IS the mount offset, so carrying the MJCF's <body name="cap" pos="0 0 Z"> into the second joint's frame as well floated the cap Z metres high.
  2. The MJCF's <contact><exclude body1="bottle" body2="cap"/> had no USD counterpart. PhysX read the authored cap-over-neck overlap as a collision and drove the cap up its own slide joint until the convex hulls cleared — cap_slide parked at 32 mm and stopped tracking the commanded value. Now emitted as UsdPhysics.FilteredPairsAPI.

Measured — cap height above the bottle body at commanded slide 0 / 18 / 36 mm (IsaacSim, states set and drawn, no physics steps: the split-render path these render-grade USDs exist for):

bottle_cap 0 mm 18 mm 36 mm
MuJoCo (source of record) 0.140 0.158 0.176
published USD 0.280 0.280 0.288
this PR 0.140 0.158 0.176

juice_shot, at 0 / 10 / 20 mm: MuJoCo 0.069 / 0.079 / 0.089, published 0.138 / 0.148 / 0.158 (exactly +69 mm, its own mount offset), this PR 0.069 / 0.079 / 0.089.

The MJCFs are unchanged, so MuJoCo behaviour and the asset physics regressions are untouched (15 passed).

Land together with WUJI_SimBench PR #124 — either alone leaves the published asset and the converter disagreeing.

Everloom changed pull request status to merged

Sign up or log in to comment