Inspect an OpenUSD scene before physics simulation

Learn in your browser · Chapter 01 of 06

Begin with a small original asset and ask what the composed stage actually contains. No GPU, rendering tool or model account is required.

An OpenUSD stage is the composed scene view, not merely the text in one file. Open the stage and inspect its prim hierarchy. In this lab, ready-candidate.usda references accessory.usda, so the marker appears under /World/Attachment even though its original definition lives in a different layer. [S1]

Use complete prim paths when identifying a finding. Shape alone is ambiguous if several branches contain a prim with that name. The project expects the moving crate body at /World/Crate and its cube at /World/Crate/Shape. A path tells the reviewer which composed object the report describes.

The default prim is an asset entry point, not a command to simulate everything below it. Our project contract expects /World as that entry point. A different valid default may serve another asset, so report a project mismatch rather than claiming that every other default prim is invalid USD. [S1]

A prim type and applied API schemas answer different questions. Cube describes geometry; PhysicsCollisionAPI adds collision-related properties. Inspect both the type and applied APIs instead of treating a Cube label as evidence that physics participation has been configured. The inspector uses IsA and HasAPI for those distinct checks. [S2]

A successful open and traversal show the composed objects currently available to that stage. They do not prove every possible variant, unloaded payload, time sample or custom resolver configuration has been checked. Our fixture path uses ordinary local references at default time; keep that limited scope in the result.

Preserve supplied source assets. The inspector reads them and returns a report without saving an edited stage. If a repair is needed, propose a separate candidate or reviewed override rather than silently modifying the source that produced the earlier evidence. Retain before-and-after identifiers so another person can trace the change.

The exercise scene is deliberately small: one crate, one floor, a material, a PhysicsScene and a local referenced marker. It is original teaching data, not a model of a real factory or a validated product. The marker is decorative; the crate and floor are the task-relevant collision objects.

from pxr import Usd
s = Usd.Stage.Open("fixtures/ready-candidate.usda")
for prim in s.Traverse():
    print(prim.GetPath(), prim.GetTypeName(), prim.GetAppliedSchemas())

Learn in your browser · Chapter 02 of 06

Scene metrics and inherited transforms can turn plausible-looking numbers into a physically different object. Inspect the result before correcting it.

Check whether metres-per-unit was authored as well as reading its value. OpenUSD returns 0.01 when that metadata is absent. The project requires an explicit value of 1.0, so a missing declaration is a finding even if an application supplies a fallback. A fallback is not evidence of the author intended unit. [S3]

Convert a fully transformed world-space length using the stage metric. In the ready fixture, the crate measures about 0.4 scene units and the metric is 1 metre per unit, giving about 0.4 metre. At 0.01 metre per unit the same world length would describe about 0.004 metre. [S3]

Ancestor transforms can change a child shape size without changing its local size attribute. The scaled-parent fixture keeps cube size at 0.4 but applies scale 100 on the crate parent. Its world dimension is therefore about 40 metres with this fixture metric, violating the project requirement. Inspect composed transforms, not just the child field. [S4]

The project explicitly uses Z as its up axis. Setting an up-axis label describes a convention; it does not rotate already authored geometry to fit that label. When combining differently oriented assets, review their actual transforms and the intended conversion rather than fixing orientation by changing metadata alone. [S5]

Transform operation order matters. Translating then scaling and scaling then translating can produce different world placement. Inspect xformOpOrder and the composed result when an AI edit moves an object unexpectedly. A list containing the same operations is not sufficient evidence that the resulting placement stayed the same. [S6]

The inspector constructs a fresh transform cache for each inspection. A reused XformCache does not automatically invalidate when stage values change, so clear or replace it before trusting post-edit results. Otherwise an apparently unchanged world position may be a stale computed value rather than proof that the edit had no effect. [S4]

Use tolerances suited to the measurement. The project allows an absolute 0.001 metre difference around each 0.4 metre crate dimension, accommodating small floating-point representation differences. It does not round 40 metres to the expected result. This tolerance is an exercise contract, not a universal engineering acceptance limit.

from pxr import Usd, UsdGeom
s = Usd.Stage.Open("fixtures/ready-candidate.usda")
print(UsdGeom.StageHasAuthoredMetersPerUnit(s))
print(UsdGeom.GetStageMetersPerUnit(s))
print(UsdGeom.GetStageUpAxis(s))

Learn in your browser · Chapter 03 of 06

Appearance, contact geometry and body motion are related but separate parts of scene description. The fixture makes those differences inspectable.

A visible object is not automatically a collider. The visual-only fixture retains the Cube and its size but removes PhysicsCollisionAPI. The scene can still describe visible geometry while failing the project collision check. Adding lights, textures or a more realistic render would not repair that missing API. [S7]

The crate Xform carries PhysicsRigidBodyAPI and the child cube carries PhysicsCollisionAPI. The first identifies the body intended to move as a unit; the second identifies collision-related shape behaviour. A collider without a rigid-body ancestor can represent static geometry such as a floor. Do not add body motion simply because an object needs contact. [S8]

Applied schema presence and enablement are separate observations. A shape may carry PhysicsCollisionAPI while physics:collisionEnabled is false. The project checks both. The API label alone is insufficient to conclude that the intended collision is enabled in the scene description. The chosen engine must still support and consume that description. [S7]

A simple proxy may represent collision more usefully than a detailed visual mesh. Its shape must still preserve task-relevant openings and contacts. For an open crate, a single filled box would erase its interior; for this solid teaching cube a box representation is appropriate. This choice is about task geometry, not visual polish.

Keep the floor static in this exercise: it has collision properties and no rigid-body API. Adding a rigid body to both objects would change the intended setup, not merely improve completeness. The reviewer must understand which objects should move before deciding whether another body API is an appropriate repair. [S8]

The PhysicsScene prim supplies scene-level physics description. Its existence is a useful structural check, not a report that a solver has run or that objects settle correctly. Our CPU inspector confirms the expected prim type without advancing time, computing contact forces or evaluating a robot controller. [S8]

The project linter is intentionally narrow. It checks the named crate collision and a scene prim, but it is not a comprehensive scan of every collider, contact filter, mesh approximation or engine extension. A passing report must state those omissions. Use an engine-specific test plan for the next gate rather than treating this fixture contract as universal readiness.

Learn in your browser · Chapter 04 of 06

A physically plausible scene needs more than a geometry label. Trace mass and material values to the object that will use them.

The lab requires a finite positive explicit crate mass and explicitly authored kilograms-per-unit. Its ready fixture uses mass 2.0 with one kilogram per mass unit. Rejecting the negative-mass fixture is our project rule. Do not infer the mass from how heavy the crate looks in a render or assume that an absent explicit mass always invalidates every USD asset. [S9]

Mass and density are not interchangeable fields. In USD Physics, a positive, non-zero explicit mass takes precedence over density on the same object; zero mass is ignored and permits density-based calculation; density-derived mass also depends on the interpreted collision volume. Recording a density value without checking which mass definition applies can give an incorrect account of the actual scene description. [S9]

A material used for appearance need not provide physics material properties. Inspect PhysicsMaterialAPI and the material bound for physics purpose. A gold-looking surface or a shader roughness setting does not establish friction or restitution. The inspector resolves the binding and then checks the material API. [S10]

A material can exist in the scene without being bound to the crate. The unbound-material fixture keeps Surface but removes its direct physics binding from the shape. Its existence in the hierarchy cannot prove that the crate uses it. Resolve the applicable bound material instead of finding any suitably named material. [S11]

Our project requires finite non-negative friction values and restitution between zero and one. The invalid-material fixture has restitution 1.2, so it fails that stated range. These numeric checks screen the authored fixture; they do not establish experimentally measured friction or predict how a particular engine combines two contacting materials.

Record the evidence behind a material value. A measured coefficient, a supplier value and an invented teaching value have different status. The lab coefficients are fictional teaching choices, not measurements of wood or metal. Do not export them as a validated library of real materials simply because their values pass a numeric range check.

Several schema getters can return fallback values when a property is not explicitly authored. Decide which fields your project requires to be explicit, and inspect authored opinions for those fields. A non-empty returned value does not by itself prove that someone measured, reviewed or intentionally supplied that value. [S2]

Learn in your browser · Chapter 05 of 06

A file is not necessarily a complete package. Keep dependencies, sources and AI-proposed changes visible to the reviewer.

A USD file can depend on other layers and assets. Package and inspect the dependency set, not just the root file. In this lab, accessory.usda is required for the composed marker. A root file that opens with a missing reference warning can still produce an incomplete stage, so opening alone is insufficient. [S12]

The inspector calls ComputeAllDependencies to report resolved layers, non-layer assets and unresolved paths. The missing-dependency fixture deliberately references absent.usda and must produce an unresolved finding. It is an expected negative fixture, not a reason to delete that report or silently replace the reference with an unrelated asset. [S12]

Use the supplied original local fixtures, not arbitrary downloaded USD files. Asset resolvers, plug-ins and dependencies can change what opening an asset accesses. This exercise is not a security sandbox or a scanner for hostile assets. Do not attach network resolvers, credentials or untrusted plug-ins merely to complete the lab.

Ask an AI assistant for a narrow proposed difference, such as restoring the crate collision API in a separate candidate. Review the actual changed fields before executing generated code. A request to make it SimReady is too vague to establish which contract, asset version or evidence the assistant should preserve.

Do not repair one finding by hiding its test. Deleting the scaled-parent check makes the report quieter while leaving the scale error intact. Preserve the expected result, make the intended asset change, reopen it and rerun the full suite. The negative fixtures help demonstrate that the checker can still notice the defect.

Keep the original composition relationship when comparing a repaired candidate. Flattening can be useful for a particular delivery, but it changes the representation and may remove the layer relationship a reviewer needs to inspect. Record any such transformation explicitly and keep the source package and provenance available. [S1]

Learn in your browser · Chapter 06 of 06

Finish with useful evidence and a bounded claim. The practical exercise is ungraded and this training is educational and non-accredited.

The executable route performs actual OpenUSD static inspection through pxr, not a handwritten USDA string search. Opening, composing, retrieving applied schemas, resolving a physics material and computing world bounds are SDK operations. It still does not execute a physics solver; name that boundary precisely in the evidence record.

The ready candidate should pass the selected project checks, and each flawed fixture should fail its corresponding check. A suite that rejects every file does not demonstrate a useful inspection process. Pair positive and negative fixtures and assert both the decision and the expected finding code rather than merely counting warning messages.

Preserve the SDK version, project-check version, root and dependency digests, fixture names and findings. The report identifies the actual bytes and environment inspected. A digest can show consistency with a reference, but cannot establish physical accuracy, author identity or lawful ownership of an asset.

A static pass is an invitation to the next test, not certification. The next owner must choose an engine, inspect collision behaviour and run an appropriate physics rehearsal. NVIDIA uses SimReady in its learning material, but this original exercise is not its validator or certification process and supplies no NVIDIA approval. [S13]

Keep authored, independently reviewed, runtime-checked and publicly released status distinct. This candidate has an actual CPU SDK inspection route; it still needs independent course/bank review, artwork, workbook/browser checks and release integration. The knowledge exam does not score the practical exercise or certify a real digital twin.

The scored final sitting contains twenty knowledge questions. Seventeen correct answers pass the strictly-above-eighty-percent rule; sixteen do not. MICKAI educational completion is non-accredited and does not establish professional competence, NVIDIA affiliation, simulation safety or learner identity.

# Run with the isolated environment described in README.md.
python inspect_scene.py ready-candidate
python inspect_scene.py visual-only
# The second command intentionally exits 1 with COLLISION_API.
python -m unittest -v test_inspector.py