fix(assets): head-heavy mixed-material hammer + recalibrated nail
fix(assets): head-heavy mixed-material hammer + recalibrated nail
Ports the same three-file fix already merged into the private
SomaStacksOrg/roboverse_data as PR #40 (merge 1795005b). This repo's main
carries the identical pre-fix content, so the diff here is byte-identical.
The YCB hammer had the right total mass in the wrong place, and fixing that
exposed that the nail's dry friction had been calibrated against the broken tool.
assets/ycb/hammer/hammer.xml — per-part densities instead of one fitted
density. A uniform 1840.5 kg/m^3 over the 7 convex parts left a 384 g handle
and a 281 g head, balancing 43% of the tool length from the striking face; a
real claw hammer carries ~4/5 of its mass in the steel head. That is not
cosmetic: a body struck off its centre of mass rotates instead of driving
through, so the effective striking mass at the nail was only 331 g of 665 g.
The split is derived from this asset's own scan texture (bare steel reads grey,
the hickory handle warm brown; interior sample points are labelled by the
nearest textured face, the handle's mass is fixed by hickory density x its true
scan volume, the head takes the published-mass remainder) -> steel head 531 g +
hickory handle 134 g = the same published 665 g, effective striking mass 599 g
(90%), balance 19% from the face. The implied head density comes out
6663 kg/m^3, 85% of solid steel, an independent check that the model
reconciles. Regenerate withpython scripts/build_ycb_assets.py --refit-mass hammer(RoboVerse).
Per-geom density is an EFFECTIVE density over the inflated convex-hull volume,
so it reads below the raw material value on purpose.
The 7 collision meshes are byte-identical to main — no geometry change.assets/local_hammer_nail/board_nail.xml — MuJoCo's frictionloss is a SOFT
constraint: a sustained load deforms it by an amount set by the constraint
impedance and NOT by the friction magnitude (sweeping frictionloss 15 -> 300 N
changed the creep by nothing). Once the corrected hammer put its real weight on
the nail, laying the tool on the nail head sank it 30 mm — past the task's
25 mm success gate with no blow struck. Adds
solimpfriction="0.999 0.9999 0.0001 0.5 2" and raises frictionloss 15 -> 60 N.
Main here also predates the 2026-07-06 solreffriction="0.004 1" fix, so that
lands in the same commit.Measured on hand_track.hammer_insert_nail (MuJoCo, tianji_wuji):
passive settle 15 s ~5 mm -> 0.09 mm
hammer laid on the nail 29.6 mm -> 5.6 mm (was an automatic success)
sustained 30 N press latched -> 6.6 mm
2 m/s blows to latch 1 -> 2
Checker still validates (visualize_handtrack_checker.py latches at frame 43);
scene soundness OK (verify_handtrack_bench_physics.py, hammer settles).assets/local_hammer_nail/board_nail.urdf — comment only. Its values
are inert (nothing integrates that joint; metasim mirrors the MuJoCo slide into
the IsaacSim render), so they stay at the old numbers rather than being split
from the usd/ that was converted from them. The comment now says so and points
at the MJCF as authoritative.
RECORDED DEMOS RE-VERIFIED (this repo only — it is the one that carries them).
trajs/parrot/grasp_hammer/tianji_wuji/trajectory.pkl.gz (50 eps) replayed through
scripts/advanced/collect_demo.py on MuJoCo/tianji_wuji, before vs after:
pre-fix uniform hammer : 38/50 succeeded, 50/50 byte-perfect snaps
fixed head-heavy hammer: 41/50 succeeded, 50/50 byte-perfect snaps
The mass change does NOT invalidate the published set — it replays slightly
better. All 50 still byte-perfect-snap because _mujoco_state.buffer is sized by
qpos/qvel, not by density, and the collision meshes are untouched. (The 38 vs 41
spread is replay sensitivity in a set that was never 50/50 to begin with, not a
regression either way.)
Known follow-ups, deliberately NOT in this PR:
- assets/ycb/hammer/usd/hammer.usd carries no mass APIs at all, so a USD consumer
computes mass from geometry at a default density, i.e. uniform. It is unused by
hand_track.hammer_insert_nail (that RigidObjCfg sets only mjcf_path, and metasim
resolves isaacsim -> usd_path/urdf_path -> None), and authoring MassAPI cannot be
verified on this box (no working IsaacSim runtime), so it is left alone rather
than shipped unverified. - usd/board_nail.usd likewise still carries friction=15 from the URDF; render-only.