fix(assets): hand_track prop physics — mass refit on 4 props + tennis-ball restitution

#39
by Everloom - opened

Five prop MJCFs, measured standalone against the shipped contact profile with
RoboVerse/scripts/audit_asset_physics.py (MuJoCo 3.8.0, each prop compiled alone over a ground
plane). Only densities and one solref change — no geometry moves, so collision shapes, grasp
affordances and resting heights are untouched.

Mass refit

prop before after target change source of the error
cups_a 0.029 kg 0.013 kg 0.013 kg density 700 → 314 absent from build_ycb_assets.py's REAL_MASS_KG, so it took the 700 kg/m³ fallback
cups_f 0.064 kg 0.021 kg 0.021 kg density 700 → 231 same fallback
feast_scented_candle 0.039 kg 0.400 kg 0.400 kg density 120 → 1219 a scented candle in a glass jar weighing 39 g
meshycup 0.677 kg 0.300 kg 0.300 kg visual geom tagged density="0"; collision 250 → 1964 see below

meshycup is the substantive one. meshycup_visual carried no density="0", so MuJoCo weighed
the render mesh at its default 1000 kg/m³ — 0.639 kg of the compiled 0.677 kg, 94 % of the mass
and the inertia, came from a geom nothing in the scene can touch.
Its sibling
meshybottle_visual was already tagged; this one was missed. Tagging it leaves only the shell, so
the collision density is refit to land a 0.30 kg glazed ceramic cup over the 0.153 L shell.

What a 10× mass error means in practice, measured: given the same 0.12 N·s impulse, the 39 g
candle slides 579 mm; the 400 g candle slides 14 mm.

Tennis-ball restitution

solref 0.01 1.00.006 0.30. Critical damping made the one prop here that is supposed to
bounce perfectly inelastic — measured e = 0.000 against the ITF spec e = 0.73 (1.35 m rebound
from 2.54 m). The new value gives e = 0.458 over 4 bounces, settling in 0.76 s.

0.73 is not reachable in this contact model at dt = 0.002. Sweeping the damping ratio down, the
ball starts gaining energy between bounces below 0.28 — measured bounce-to-bounce e = 1.32 and
1.63, which is numerical, not a springier ball. 0.30 is the stiffest setting whose successive
bounces stay physically consistent (0.40, 0.46).

This one does not take effect on its own. HAND_TRACK_SIM_PARAMS overwrites every geom's
solref in the assembled scene, so the authored value is clobbered at load — re-measured with the
override on, e is back to 0.000. It needs a backend change (skip geoms that author their own
solref, or an explicit per-asset opt-out). The asset is correct now; the pipeline still discards
it. Landing this file is still right — it stops the asset from being wrong once the backend allows
it through.

Land together with the code side

cups_a / cups_f must be added to REAL_MASS_KG in RoboVerse/scripts/build_ycb_assets.py, or
the next rebuild silently restores the 700 kg/m³ fallback and reverts both files. That is a
separate change in the code repo and should land alongside this PR.

Caveats for reviewers

  • Every mass change is a re-collection event. Demos on disk were recorded under the old masses
    and PD action-replay reproduces recorded torques, so the replayed trajectory changes with the
    mass. feast_scented_candle's density of 120 was explicitly copied from the Parrot scene so
    replay reproduced the recorded contact dynamics — this does shift pack_box and
    place_candle_in_box action→motion. Snapped-state replay is unaffected. Re-run the grasp
    regression on nest_cups, pour_bottle_cup, pack_box, place_candle_in_box, grasp_candle
    before trusting any success rate measured against these props.
  • feast_soda_can is NOT in this PR and carries the identical defect — the same density 120,
    compiling to 0.042 kg for a full can. It was left out because it was not on the fix list; flagging
    in case that was an oversight.
  • meshybottle is deliberately unchanged. An earlier pass called it 2.2× too heavy by judging
    1.100 kg against a 0.5 kg PET water bottle. The MJCF documents a deliberate 2026-07-05 rebalance
    to a partially-full wine bottle — explicit <inertial>, COM dropped for the liquid. 750 ml of
    wine is ~0.75 kg before the glass, so 1.1 kg is right. The reference was wrong, not the asset.
  • Reference masses for cups_a / cups_f are the least certain numbers here (the YCB 065 stacking
    cups) — worth a check against the published listing before merge.
  • meshycup.xml ships CRLF line endings; they are preserved.

Code side is up: somastacks/WUJI_SimBench#129 — these two must land together.

That PR adds cups_a / cups_f to REAL_MASS_KG in build_ycb_assets.py. Without it the next asset rebuild recomputes both cups from the 700 kg/m³ fallback and silently reverts the two files in this PR.

It also carries the guard for the ---in-XML-comment breakage that hit this PR mid-review: a backend-free test that strict-parses every asset MJCF, because MuJoCo'''s TinyXML2 tolerates the illegal token and will not tell you. All 74 asset XMLs parse with ElementTree now, and the guard was checked in both directions.

Everloom changed pull request status to merged

Sign up or log in to comment