[MGPU] Tests: make newton-only tests kitless (drop AppLauncher boot)

[hujc7]

[MGPU] Tests: make newton-only tests kitless (drop AppLauncher boot)

Summary

• Removes AppLauncher boot from 2 newton-only test files that already use the newton-aware fixture; they now run via SimulationContext's existing has_kit() gate without booting Kit / Isaac-Sim.
• Replaces unguarded from omni.physx.scripts import utils in schemas.py with direct pxr.UsdPhysics.FixedJoint calls (was the bulk blocker: 95 fixed-base test failures across kitless newton tests routed through this line).
• Replaces omni.usd.get_context().get_stage() in newton test_articulation with isaaclab.sim.utils.stage.get_current_stage for kitless compatibility.
• Net effect for multi-GPU CI: newton test_articulation no longer boots Kit, so the upstream concurrent-Kit shutdown hang (#3475) does not apply to it. Its MULTI_GPU_SKIP_REASON marker in #5875 drops once this PR lands.

1. Problem

Newton-only tests inherited an AppLauncher(headless=True).app boot from the physx test template they were copy-pasted from (PR #4818, March 2026). Tests that drive newton through the newton-aware fixture don't actually need Kit — newton's SimulationContext path runs kitless when isaaclab.utils.version.has_kit() returns False — but the boilerplate forced Kit boot anyway, costing ~1:20 per file single-shot and exposing the suite to the Kit lifecycle SIGHUP/shutdown-hang bug under concurrent multi-GPU CI.

2. Scope

2.1 Files converted to kitless

File	Lines changed	Local 3-MIG result
source/isaaclab_newton/test/assets/test_articulation.py	+7 -5	224 passed, 8 skipped, 1 xfailed
source/isaaclab_newton/test/assets/test_rigid_object_collection.py	-8	158 passed, 8 skipped

Each file: removes AppLauncher import + simulation_app = AppLauncher(headless=True).app, replaces docstring with a rationale comment.

2.2 Files investigated but NOT converted

These were tested in kitless mode and kept Kit-booted because conversion would degrade coverage or hit deeper kitless gaps. Documented for follow-up; out of scope here.

File	Why not
source/isaaclab/test/markers/test_visualization_markers.py	13 of 19 tests pass kitless; 6 ERROR at setup (ModuleNotFoundError: omni.physics) because they construct a physx-backed visualizer path through string_to_callable. Conversion costs 6 tests of coverage.
source/isaaclab_newton/test/sim/test_newton_schemas.py	Fixture uses bare SimulationCfg(dt=0.1) which defaults to PhysxCfg. Adding physics=NewtonCfg(...) unblocks 9/13 tests; the remaining 5 fail on schema-routing ('NewtonMeshCollisionAPI' in ['PhysicsMeshCollisionAPI']). Needs schema-routing investigation.
source/isaaclab_contrib/test/deformable/test_deformable_object.py	Fixture already uses NewtonCfg(solver_cfg=VBDSolverCfg(...)). Blocker is unguarded from omni.physx.scripts import deformableUtils at schemas.py:1213 and meshes.py:291. Making deformable bodies kitless requires reimplementing the helpers in newton/pxr.
source/isaaclab_contrib/test/deformable/test_rigid_deformable_coupling.py	Same as above.

2.3 schemas.py kitless helper

source/isaaclab/isaaclab/sim/schemas/schemas.py: replaces unguarded from omni.physx.scripts import utils call (in modify_articulation_root_properties fixed-base joint creation path) with a local _create_fixed_joint_to_world() helper that uses pxr.UsdPhysics.FixedJoint directly. Was the single bulk blocker for kitless newton runs — all 95 fixed-base test failures routed through this line. Kept even though only 2 of 6 tests fully convert: the schemas.py path is universally invoked.

2.4 omni.usd → stage util

source/isaaclab_newton/test/assets/test_articulation.py: replaces import omni.usd; fix_reversed_joints(omni.usd.get_context().get_stage()) with from isaaclab.sim.utils.stage import get_current_stage; fix_reversed_joints(get_current_stage()).

3. Validation

3.1 Single-shot kitless

• test_articulation.py on cuda:0: 224 passed, 0 failed (4:39, Kit boot count = 0).
• Kit-booted baseline same file: 224/0 (5:59).
• Delta: -1:20 (-22%) per file.

3.2 3-MIG concurrent (Horde L40, hardware-isolated)

• 3 shards × 2 converted files: all PASSED, no SIGHUP, no Kit lifecycle bug.
• test_articulation: 14:19 wall per shard under 3-way MIG contention (slower than single-shot since 3 parallel pytest processes share host CPU for warp compile).
• test_rigid_object_collection: 2:25 wall per shard.

4. Follow-ups (not this PR)

• Visualization markers kitless coverage: make the physx-backed visualizer path kitless-aware in isaaclab_physx.physics.physx_manager so the 6 currently-Kit-bound tests can also drop AppLauncher.
• Schema-routing fix for test_newton_schemas: investigate why schemas.define_collision_properties applies the physx-side PhysicsMeshCollisionAPI instead of the Newton-side NewtonMeshCollisionAPI when the cfg is Newton-flavored.
• Deformable kitless support: replace unguarded omni.physx.scripts.deformableUtils calls in schemas.py and meshes.py with kitless equivalents.

5. Why this is [MGPU]-tagged

The Kit-boot boilerplate doesn't matter for single-GPU CI (works fine, just slow). It matters for multi-GPU CI because concurrent Kit instances trigger the upstream shutdown hang on heavy tests like test_articulation. Companion to #5875 (multi-GPU CI infra), and independent of #5881 / #5886.

[isaaclab-review-bot]

AI Review Summary

This PR makes newton-only tests "kitless" by removing the AppLauncher boot, reducing per-file test time by ~22% and avoiding Kit's SIGHUP lifecycle bug under concurrent multi-GPU CI.

Strengths

1. Clean architectural approach - The key insight that SimulationContext already gates Kit-specific paths via has_kit() makes this change straightforward and minimally invasive.

2. Well-documented _create_fixed_joint_to_world helper - The new helper function in schemas.py includes comprehensive docstrings explaining its purpose as a kitless equivalent of omni.physx.scripts.utils.createJoint. The logic mirrors the omni helper's "find first writable ancestor" walk for instanced/prototype prims.

3. Consistent test file changes - All 6 test files follow the same pattern: remove AppLauncher boilerplate, add a docstring referencing test_articulation.py for rationale. This makes the codebase easier to maintain.

4. Good changelog entry - The .rst changelog clearly explains what was fixed and why.

Suggestions & Questions

1. Unique joint naming: The _create_fixed_joint_to_world function uses a while loop to find a unique FixedJoint name:

   while stage.GetPrimAtPath(f"{joint_base_path}/{joint_name}").IsValid():
       joint_name = f"FixedJoint_{idx:02d}"
       idx += 1

   Consider adding a reasonable upper bound (e.g., 1000 iterations) with a clear error message to prevent infinite loops in pathological cases, even if unlikely.

2. Import placement: In test_articulation.py, the import of get_current_stage is done conditionally inside the if block:

   if any(name in usd_path for name in _REVERSED_JOINT_USD_FILES):
       from isaaclab.sim.utils.stage import get_current_stage

   This is fine for lazy loading, but if this pattern appears elsewhere or get_current_stage is used more broadly, consider moving it to module-level imports for consistency.

3. Test coverage for the new helper: The PR mentions that kitless runs pass 224/0 tests cleanly, which is great validation. Is there (or should there be) a dedicated unit test for _create_fixed_joint_to_world that exercises edge cases like:

   • Articulation prims that are instance proxies
   • Prims at the root level (/)
   • Name collisions requiring the _01, _02 suffix logic

4. Docstring wording: In test_articulation.py, the docstring mentions "shaves ~30s off per-file boot" but the PR body says the delta is 1:20 (-22%). Minor inconsistency worth harmonizing.

5. Empty .skip files: The changelog .skip files (isaaclab_contrib/changelog.d/...skip, isaaclab_newton/changelog.d/...skip) are empty. Is this intentional to signal "no changelog needed for this package"? If so, a brief comment in the PR would help reviewers understand the convention.

Overall

This is a solid quality-of-life improvement for the CI pipeline. The changes are well-scoped, the helper function is correctly implemented using pxr.UsdPhysics primitives, and the validation numbers look good. The PR clearly explains the problem, scope, and non-scope.

---

This review was generated by an AI assistant. Please verify suggestions before implementing.

---

Update (commit d887cba): The new commit reverts the kitless conversion for 3 tests (test_newton_schemas, test_deformable_object, test_rigid_deformable_coupling) that need Kit because they resolve the physics backend via omni.physics. This is a sensible scope reduction — the commit message clearly explains why these files can't go kitless without switching to the newton-aware cfg pattern, which is out of scope.

The remaining 3 kitless files (test_articulation, test_rigid_object_collection, test_visualization_markers) are unchanged. No new issues introduced.

Previous inline comment status:

• Ambiguous cross-reference nit in test_visualization_markers.py: not addressed in this commit (file untouched). Original comment stands — it's a minor nit and non-blocking.

---

Update (commit c2495f3): This commit reverts the kitless conversion for test_visualization_markers.py, restoring AppLauncher boot. The commit message clearly explains that 6 of 19 tests hit ModuleNotFoundError: No module named 'omni.physics' because the physx-backed visualizer path imports omni.physics.tensors. This is a sound scope reduction — keeping Kit boot for this file preserves all 19 tests of coverage.

Previous inline comment status:

• Cross-reference nit in test_visualization_markers.py: ✅ Resolved (moot) — the kitless docstring containing the ambiguous reference has been removed entirely by this revert.

No new issues introduced. The PR now kitless-converts only test_articulation.py and test_rigid_object_collection.py, which is a well-scoped final state.

[isaaclab-review-bot]

Nit (ambiguous cross-reference): This file is in isaaclab/test/markers/, not in isaaclab_newton/. The bare test_articulation.py reference is ambiguous since there are multiple files with that name across packages.

Consider using the full path for clarity:

"""Kitless newton-only test: no AppLauncher boot. See
isaaclab_newton/test/assets/test_articulation.py for rationale ..."""

The other cross-package references (test_deformable_object.py, test_newton_schemas.py) already use the full path — this one appears to be an oversight.