sample_content: obs_workbench_tool_a01 worktop faces are inverted — fails its own culling_xz benchmark (27.8% vs 10% threshold)

[NV-JBapst]

Summary

The sample prop obs_workbench_tool_a01 has an inverted top surface. Under back-face culling the entire worktop disappears and the unlit interior of the cabinet shows through. It fails this repository's own culling_xz FET001 benchmark.

This is the reference asset shipped for the benchmark that detects the defect.

Reproduction

simready-validate ... --stamp-asset-validation sm_obs_workbench_tool_a01_01.usd
simready-benchmark --assets sm_obs_workbench_tool_a01_01.usd --max-concurrent 1

Asset: sample_content/common_assets/props_general/obs_workbench_tool_a01/isaacsim_usd/

Result:

FAIL  culling_xz   XZ-axis culling check failed at angle 0 deg:
                   27.8% object pixel diff (threshold 10%).

The other seven tests in the run pass (ground_drop, slope_drop, light_response, normals_xz, pivot, presence), so rendering and framing are working — this is not the "no frames captured" or "camera misframed" failure mode from the test doc.

Visually: with doubleSided = True the asset renders correctly in orange diagnostic material. With culling on, the worktop renders solid black while the drawer fronts, sides and edges remain correct. Happy to attach the captured frame pair if useful — culling_xz_worst_pair.mp4 is produced by the run.

Root cause

Offending prim:

/sm_obs_workbench_tool_a01_01/workbench_main_obj/workbench_main_trans/workbench_main_obj/workbench_main_mesh

Geometry lives in payloads/obs_workbench_tool_01_meshes.usd.

Checking face orientation on that mesh (stage upAxis = Z, mesh Z range -0.0010 .. 0.9134):

Surface	normals UP	normals DOWN
Top faces (z ≈ 0.9134)	0	35
Bottom faces (z ≈ -0.0010)	0	4

Every one of the 35 faces forming the worktop points downward, into the object. The bottom faces also point down, which is correct for a bottom surface — so the top surface is simply inverted relative to the rest of the model.

Worth noting: this is not inconsistent winding

The test documentation attributes a culling_xz failure to "incorrect (clockwise) winding order". That is not what is happening here. Comparing winding-derived normals against the authored normals attribute, every face of every mesh in the asset agrees — 496/496 on workbench_main_mesh, 403/403 on each of the ten drawers, 0 opposed, all faceVarying.

Winding and normals are consistent with each other; they are jointly inverted on the top surface. So the fix is not "flip the winding to match the normals" — both need flipping together (or normals recalculated outward after the flip). The culling_xz "How to Fix" guidance may be worth widening to cover this case, since following it literally would produce a mesh whose winding and normals disagree.

doubleSided is authored True on the mesh, which is why the defect is invisible in most DCC viewports and in any renderer that does not cull.

Scope not checked

Only the isaacsim_usd variant was tested. The asset also ships simready_usd and simready_physx_usd variants plus the original Blender source at dcc_source/working/model/blender/obs_workbench_tool_a01.blend — worth checking whether the inversion originates upstream of the USD conversion and affects the other variants too.

Environment

simready-benchmark 2026.6.5, simready-validate 2026.6.4, Isaac Sim 6.0.1, driver 596.86, RTX 6000 Ada, Windows 11, Python 3.12. Repo at 0ed0dfb.

[NV-JBapst]

This is not isolated — 3 of the 14 sample props fail culling_xz, and the workbench case is a conversion regression

I swept the rest of sample_content/common_assets/props_general (27 asset variants across 14 props). Two findings that change the shape of this report.

1. Two more props fail, and worse than the workbench

Asset	Variant	culling_xz result
obs_lamp_revolute_a01	simready_usd	FAIL — 63.2% object pixel diff at 180°
obs_joystick_a01	simready_usd	FAIL — 39.7% object pixel diff at 180°
obs_workbench_tool_a01	isaacsim_usd	FAIL — 27.8% at 0° (this issue)
apple_a01	simready_usd	PASS (control)
obs_small_sledge_hammer_a01	simready_usd	PASS (control)

Both new failures are on base_mesh_00 and both fail at 180° — the underside — where a static check finds inverted faces on the lower shell (obs_lamp_revolute_a01: 681 inward vs 688 outward faces; obs_joystick_a01: 129 inward). Unlike the workbench, both fail in both their simready_usd and simready_physx_usd variants, so these look inherited from the source model rather than introduced downstream.

2. The workbench defect appears to be introduced by the Isaac conversion, not present in the source

This is the more interesting result:

obs_workbench_tool_a01 variant	culling_xz
simready_usd	PASS
isaacsim_usd	FAIL — 27.8%

The same prop passes in its simready_usd form and fails in its isaacsim_usd form. A static comparison agrees: in isaacsim_usd, workbench_main_mesh has 133 inward-facing bounding-box-shell faces against 4 outward, spread across five directions; in simready_usd the main body does not flag at all — only the drawers show a small, benign hit that the runtime test passes.

If that holds up, the inversion is being introduced somewhere in the Isaac conversion path (metadata.json records PropAssetHandler v2.0.0, 2026-07-17) rather than existing in the authored art. That would make it a pipeline bug affecting anything converted that way, which is a wider blast radius than one bad asset. Happy to split it into its own issue if you'd prefer to track it separately.

Note obs_electricians_large_tool_box_a01/isaacsim_usd — the only other isaacsim_usd variant in the sample set — produced no meshes when traversed from the main asset file, so it was not comparable here. Might be worth a look on its own.

Method

Two independent checks, agreeing:

1. Static screen — on each mesh, faces lying on the extremes of the mesh bounding box are unambiguously exterior, so their geometric normal must point away from the object. Counting those that point inward flags candidates in seconds. This has false positives by construction (a mesh with no modelled bottom face flags on its underside), which is why every hit was confirmed with the runtime test. Signed volume via the divergence theorem is not usable here — none of these meshes are closed (the workbench body has 32 boundary edges; the drawers are non-manifold).
2. simready-benchmark --tests culling_xz — ground truth.

Every strong static hit failed the runtime test; the one weak hit passed; both clean controls passed.

Environment as in the original report: simready-benchmark 2026.6.5, Isaac Sim 6.0.1, repo at 0ed0dfb.