cabinet_12: one collision decomposition for the three identical bins

#24
by Everloom - opened
SomaStacks Inc org

What

cabinet_12's three tilt-out bins are the same moulded part — byte-identical visual meshes,
identical extents (0.1084 x 0.32 x 0.4208 m) — yet main ships two different collision
representations for them: E_drawer1_4 carries 36 CoACD convex hulls, while E_drawer1_1 and
E_drawer1_7 carry a 50-box cavity_repair_voxbox proxy.

This PR gives all three bins the same colliders: the CoACD hull set the factory already produced
for this part, applied to E_drawer1_1 and E_drawer1_7 in MJCF, URDF and USD so every
backend sees the same geometry.

Why

The bins' cavity_repair.json records show the gate that picks the representation — CoACD is kept
unless the voxel-box proxy leaves more of the bin interior free:

bin revision CoACD free voxbox free verdict
E_drawer1_7 710e52db 76.7 % 71.0 % rejected: gate not met -> hulls kept
E_drawer1_1 710e52db 61.3 % 69.9 % accepted -> voxbox
E_drawer1_7 main 64.2 % 71.1 % accepted -> voxbox

The same part scored 61.3 / 64.2 / 76.7 % across three CoACD runs, so which representation a bin
ends up with is decided by decomposition noise rather than by the part. That is also why main
regressed E_drawer1_7 from hulls to boxes.

Measured effect (MuJoCo, colliders vs the part's own visual shell)

metric main (bins 1 & 7) this PR (all three bins)
collider material outside the visual shell 14.65 mm 0.00 mm
visible surface with no collider behind it (p99) 4.00 mm 0.04 - 0.69 mm
bin interior occupied by colliders 1.8 - 2.3 % 0.0 - 0.2 %
probes seeded in the bin and settled (hold test) 12 / 12 12 / 12
max visual/collision centre delta (bundle) 3.44 mm 0.47 mm
total mass 7.0000 kg 7.0000 kg (unchanged)

The 14.65 mm overshoot is material that a hand collides with outside anything visible; with all
three bins open the overshooting boxes on neighbouring bins wedge fingertips at the corners. The
hold test seeds probe objects inside an opened bin and settles them, i.e. it checks the
pass-through failure the voxbox policy exists to prevent: neither representation loses a probe.

Provenance — please read before pinning

The USD collision scopes were copied prim-for-prim (Sdf.CopySpec) from revision
710e52db2bb581ede393d2eaea6cff98ce20ee8c's E_drawer1_7 scope, so nothing here is
hand-authored geometry; the result is 36 Mesh prims with physics:approximation = convexHull
per bin, matching how E_drawer1_4 is authored in the same file.

usd.canonical_usd_sha256 is recomputed (that field is sha256 of the .usd bytes — verified
against the current revision). source_sha256, properties_sha256 and
dependency_closure_sha256 come out of the factory's USD canonicalisation, which is not
reproducible outside it, so they still describe the previous USD. To keep consumers from
trusting stale provenance, usd.status is set to regenerated_pending_factory — the acceptance
suite fails closed on it by design. The asset factory should re-run over this bundle and
restore status: generated with real fingerprints before anyone pins this revision.

Verification

  • MuJoCo: three-direction audit (outside-the-shell / uncovered-surface / blocked-cavity) plus the
    hold test above; masses and base collision bounds unchanged.
  • Acceptance suite test_hand_track_articulated.py against a bundle root containing this
    cabinet_12: 56 passed, 1 failed, the single failure being the deliberate usd.status flag.
  • USD checked structurally (prim types, counts, approximation attr, point extents identical
    across the three bins). Not verified: a PhysX runtime load in IsaacSim.
  • Consuming spec deltas for the downstream code pin (ArtVIPTabletopSpec.cabinet_12):
    target_sweep_bounds_local_xy ((-0.479559, 0.198729), (-0.317850, 0.235888)) ->
    ((-0.472691, 0.192694), (-0.313360, 0.222057));
    target_sweep_min_z_after_seating_m 0.282877 -> 0.286877;
    base_collision_bounds_local_xy and total_mass_kg unchanged.

Follow-up (upstream, not in this PR)

Make the decomposition deterministic per part, or decompose a part once and reuse it across
instances, so identical parts stop getting different collision policies across bins and revisions.

Everloom changed pull request status to merged

Sign up or log in to comment