[AntoineRichard]

Description

The asset tests repeated expensive scenes, replayed the same contracts on every backend, and carried custom scaffolding that made them hard to follow. This PR consolidates them across PhysX, Newton, and OVPhysX into lean, directly readable tests: cheap mocked contracts own the shared API, and real-solver tests only cover what needs the real solver.

This is the combined PR for review and merge. The split PRs are not being updated.

Test organization

• One file per asset family and backend. Every test file pays a fixed 4–7 s of imports (and a Kit launch for PhysX and in the CI core job), so tests live in the file that already pays those costs: unit-level checks sit beside the real-solver tests of the same asset, rigid objects and rigid-object collections share one file and one scene, and the shared contracts are two files. The asset tests went from 39 files to 19.
• Shared contracts (source/isaaclab/test/assets/contract/): test_articulation_contract.py (including joint/body ordering) and test_rigid_object_contract.py (rigid objects and collections, one factory) run every backend on mocked views. backends.py owns the Kit/kitless bootstrap, backend availability, and the manager doubles; unavailable backends collect as skips. Newton runs on CPU only, since its asset code has no CUDA-specific path; PhysX and OVPhysX keep CUDA rows for their pinned-staging paths.
• Writers are one table-driven test per family covering warp and torch inputs, index and mask selection, partial writes through the public getters (unselected entries must be untouched), scalar writes, and oversized-input rejection.
• Ordering is a small set of table-driven contracts (property routing, body-state reads, duplicate selectors, stale staging, reset rereads) using cyclic orderings, since a reversal is its own inverse and cannot catch a swapped index map.
• Backend tests reuse module-scoped composite scenes, with two environments so partial writes prove untouched state, and build each scene on one device unless a Lab code path differs by device (OVPhysX CPU stage serialization, CUDA-graph capture, CPU-only regressions). Tests that own a simulation run before the composite scene; PhysX asserts this so a misplaced test fails loudly.
• Actuator equivalence on PhysX and Newton runs on the local two-leg and cartpole USD fixtures (in isaaclab_newton/test/assets/data) instead of ANYmal.
• WrenchComposer keeps literal unit cases and one real PhysX delivery test. Newton OSC controllers run closed-loop tracking and gravity compensation on an island of the Newton articulation scene.

What was removed or merged

• Custom framework: capabilities.py, contract_boot.py, mock_backends.py, public_surface.py, and tests that only exercised test scaffolding or a copied member inventory.
• Duplicate index/mask, root/body/joint/tendon, and rigid/collection writer tests; per-helper _make_* variants; tests that rebuilt the same mocked asset to check overlapping properties.
• PhysX: Franka and Shadow Hand scenes (local branching fixture covers the Jacobian, gravity compensation, and fixed tendons), the 800-step IK test, ~15 ANYmal actuator simulations reduced to 3, and gravity tests folded into the shared scenes.
• OVPhysX: unit files merged into the asset files that already pay the same imports; collection tests share the rigid-object scene; heterogeneous clone tests share scenes; collision-filter and rejection variants no longer doubled across devices; reset checks folded into the wrench tests.
• Newton: rigid-object scenes run on CPU only (CUDA stays covered by the actuator, CUDA-graph, and capture tests); authoring tests merged into one table; shorter controller tracking loops; eight files merged into three.
• A source-inspection WrenchComposer test, and redundant dtype/selector matrices.
• device_split runner support (tools/_device_split.py, its test, the pytest marker, and workflow mentions). No test declares it after the OVPhysX device fix below.
• DC_MOTOR_ACTUATORS and ActuatorStateResetBase from isaaclab.test.utils.actuator_equivalence, which no longer have callers.

Every merged or removed contract has a keeper. Non-obvious merges were checked with a one-line production mutation that turns the keeper red (for example, writer kernels ignoring the mask, COM caches not invalidated, Jacobian link shift skipped, collection COM/inertia fixes reverted).

Coverage deliberately given up: the Lab-path DelayedPD reset check now runs on Newton only (PhysX asserts reset forwards environment IDs); armature-to-mass-matrix invalidation is no longer tested because PhysX does not reflect armature writes in the mass matrix within one step; PhysX composite-scene checks run on CUDA only (CPU staging stays covered by the mocked contracts); OVPhysX collision isolation, mimic fingers, and non-geometry heterogeneous articulation variants run on CUDA only.

Runtime changes included in this PR

• Fix PhysX rigid-object-collection COM and inertia setters (stacked from Fix PhysX rigid-object collection COM and inertia setters #8131).
• Allow OVPhysX CPU and CUDA simulations in one process: configure each USD physics scene for the selected device instead of pinning the process to its first device or using the sticky CPU mode. Lifecycle coverage includes CPU → CUDA → CPU.

Validation

Each file ran in its own process on an NVIDIA RTX PRO 6000 Blackwell (warm Warp cache), comparing the head of this PR before the cleanup (ad9b10a37) with the current one:

Scope	Files	Time before → after
Shared contracts	5 → 2	40 s → 17 s
Core WrenchComposer, actuators, terminations	5 → 4	19 s → 15 s
Newton assets, actuators, controllers	8 → 3	84 s → 42 s
PhysX assets	9 → 5	85 s → 40 s
OVPhysX assets and physics manager	12 → 5	100 s → 52 s
Asset tests total	39 → 19	328 s → 166 s

Current results: 1,042 passed, 31 skipped, 10 strict xfails, 99 deselected (Newton CUDA contract rows). The OVPhysX sensor, view, and randomization files are unchanged apart from dropping their device-split fixtures (~134 s, not counted above). Fewer files also means fewer Kit launches in CI, where each one costs more than locally.

Formatting, lint, LFS, and whitespace hooks pass. The changelog check passes against upstream develop.

Known failures pinned as strict xfails

• OVPhysX partial body writes overwrite unselected bodies of the selected environment (rigid-object collection).
• Newton fixed-tendon mask setters accept data with extra environments.
• PhysX and OVPhysX rigid objects do not reject an enabled articulation root.
• After a failed PhysX initialization, the first CPU simulation reads zero poses for kinematic collection bodies. The previous file order masked this.
• GPU PhysX applies mass and inertia writes to the mass matrix and gravity compensation on the next step.

Other limitations exposed by the redesign and left out of scope: full PhysX collection reset leaves wrench composers active; GPU link-frame wrench delivery can use stale transforms after teleports; Newton native-actuator reset is model-wide.

Type of change

• Test-suite redesign and cleanup
• Bug fixes for PhysX collection setters and OVPhysX device lifecycle
• CI tooling cleanup (device_split removal)

Checklist

• Read the contribution guidelines
• Run the formatting and lint checks
• Apply the test-audit criteria and verify keepers with mutations
• Validate the changed tests on CPU and CUDA where configured
• Add changelog fragments for every touched package
• Make corresponding documentation changes for the runtime behavior
• Existing contributor listed in CONTRIBUTORS.md

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Reorganizes and expands asset contract tests across physics backends.

No confirmed new issue from the readability cleanup blocks merging.

Summary

The PR consolidates cross-backend asset contracts, simplifies backend tests, adds local Newton USD fixtures, and includes PhysX collection-setter and OVPhysX device-lifecycle fixes. The latest readability cleanup yielded no confirmed new finding.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  T[Directly collected asset contracts] --> F[Backend and device fixtures]
  F --> M[Mock asset factories]
  M --> P[PhysX]
  M --> N[Newton]
  M --> O[OVPhysX]

Loading

_{Reviews (2) · Last reviewed commit: "Simplify asset tests and remove custom t..."}

[isaaclab-review-bot]

Isaac Lab Review Bot

The asset-test redesign establishes shared contracts, backend-focused units, and small real-solver seams, with accompanying PhysX and OVPhysX fixes. Before merge, remove the committed benchmark/work-report artifact and eliminate the redundant, unenforced public-member snapshots.

• Design and architecture: The contract/unit/integration ownership split and explicit backend-capability table are coherent. However, the committed redesign-results document is a machine- and revision-specific work artifact prohibited by repository guidance, and public_surface.py contains unused duplicate member inventories that can silently diverge from the enforced mappings.
• API: The touched production APIs retain their signatures: PhysX collection COM/inertia changes reshape only at the TensorAPI boundary, deformable type inference preserves material-before-topology precedence, and the OVPhysX device lifecycle change affects internal manager behavior. The public-surface audit is useful, but its unused duplicate snapshots should be removed or made authoritative.
• Implementation: The contract bootstrap, backend capability declarations, manager patch scoping, writer readbacks, fused-layout checks, and backend regression seams were traced. Required cleanup is limited to removing the generated benchmark report from the repository and consolidating the duplicated public-member inventory.

Minor fixes needed. Posted 2 actionable findings inline.

Automated review; human maintainers own approval decisions.

[isaaclab-review-bot]

🟡 Warning · Design Architecture — Committed generated benchmark results artifact

This file is a generated work report: it contains a "Copy-ready PR performance section", a rebase-specific base SHA, a machine-specific environment (GPU, driver, wheel versions), sandbox paths such as /tmp/isaaclab-task8-warp, and a per-module disposition map. Repository guidelines prohibit committing generated plans, scratch files, or agent artifacts. It will go stale immediately after merge and carries no reader-facing documentation value. Keep these measurements in the pull-request description instead.

[isaaclab-review-bot]

🔵 Suggestion · Design Architecture — Unused member snapshots duplicate enforced mapping

Only the BaseRigidObjectData, BaseRigidObjectCollectionData, and BaseArticulationData entries of _PUBLIC_MEMBER_SNAPSHOT feed PUBLIC_SURFACE_CONTRACTS; the AssetBase, BaseRigidObject, BaseRigidObjectCollection, and BaseArticulation entries are never referenced and respell the same member names that the _mapping(...) calls list inline. A future API change updates only the audited copy, leaving a silently diverging second inventory. Drop the unreferenced entries or build the mappings from them.

[AntoineRichard]

Consolidated to 13 PRs

The series is now 13 PRs instead of 24: #8134, #8135, #8138, #8139, #8141, #8144, #8146, #8147, #8149, #8151, #8153, #8154 and #8156, titled "Accelerate and strengthen asset tests [N/13]".

• What was merged: parts that edited the same file (the articulation halves) and small owner PRs whose coverage they received now share one PR. The 11 absorbed PRs are closed, each pointing to the PR that absorbed it. No code changed in the consolidation.
• Review round: before consolidating, every PR went through an independent review with mutation checks. The coverage gaps it found are closed, each proven by the reviewer's own mutation.
• Other fixes from that round:
  • Accelerate and strengthen asset tests [3/13] #8138 now sets the device on every physics scene in the stage.
  • Eight OVPhysX test files no longer skip the second device.
  • The WrenchComposer delivery test runs on CPU again.
  • isaacsim_ci routing matches develop.
• Timing: measured serially on one machine, the five asset scopes take 374.9 s → 150.5 s (2.5×). The description has the full table, sizes and dependencies.

[StafaH]

run-ci