[CI] Stabilize the Newton environment smoke tests and run them kit-less

[hujc7]

Description

The isaaclab_tasks shards intermittently aborted with SIGSEGV or SIGABRT while creating a
Newton environment — measured at ~5% per environment build. The abort is glibc catching a
corrupted heap (malloc(): unaligned tcache chunk detected) at an unrelated later allocation, so
pytest recorded CRASHED with no Python traceback and no failing node.

Root cause is a race in OpenUSD's physics parser: collecting collider descriptors for several
colliders under one rigid body reallocates a shared std::vector across worker threads.

"Fixed a bug in UsdPhysicsParsingUtility which could cause a race condition and crash when parsing
multiple colliders under the same rigid body in a multi-threaded context" — OpenUSD CHANGELOG
[26.05] - 2026-04-24, upstream issue 4002

This PR contains three related changes.

1. Work around the parser race

Limit OpenUSD's work-thread pool for this suite, mirroring the workaround already in
source/isaaclab_tasks/test/rendering_test_utils.py and carrying the same removal note.

Applied at module scope rather than in a fixture: kit-less, OpenUSD builds its task arena while
tests are collected, and WorkSetConcurrencyLimit does not shrink an arena that already exists
(measured — it still reported 128 workers and still crashed).

2. Run the suite kit-less

The MJWarp preset is a kit-less backend, so AppLauncher is dropped and the suite filters to tasks
that run without Kit. The split is derived from scan(...).needs_kit — the same decision
launch_simulation() makes — so a task follows its own config instead of a hand-maintained list.

Tasks carrying Kit camera sensors are excluded; test_environments.py already smoke-tests each of
them on its default backend, so no task loses coverage.

get_isaac_sim_version() in _run_environments is now called only when create_stage_in_memory is
set, since it imports isaacsim — that unconditional call was what previously forced Kit.

3. Make the physics preset selection actually apply

Selecting a preset in this suite had no effect. parse_env_cfg resolves every PresetCfg wrapper to
its default before returning, and the follow-up apply_overrides call had no choice node left to
act on, so each task ran whichever backend it defaulted to.

parse_env_cfg now takes the selection and applies it while the wrappers are resolved, which is
what register_task_to_hydra already does. Default presets=() keeps every existing caller
unchanged. The raw-config / collect_presets / apply_overrides dance in the test util collapses to
a single parse_env_cfg(...) call.

Consequence: Isaac-Velocity-Flat-G1 and Isaac-Velocity-Flat-Spot now run Newton in this
suite rather than PhysX. Both pass.

Evidence

Standalone reproducer — pxr only, no Isaac Lab, Newton, CUDA or Kit; 15 processes x 40 parses of
the Franka Reach stage:

usd-core	Has the 26.05 fix	Crashed processes
25.11 (current pin)	no	14 / 15
26.3	no	10 / 15
26.8	yes	0 / 15

Crash rate against OpenUSD's thread count, which is what this PR limits:

PXR_WORK_THREAD_LIMIT	1	2	4	8	16	32	128
crashed procs (of 15)	0	0	10	10	12	11	14

Captured stack, previously unavailable because the process died without a Python traceback:

#0  moveDescsToDict<UsdPhysicsRigidBodyDesc>(...)        pxr/UsdPhysics/_usdPhysics.so
#1  LoadUsdPhysicsFromRange(stage, paths, callback, ...) libusd_ms
    newton/_src/utils/import_usd.py:546  parse_usd
    isaaclab_newton/cloner/newton_clone_utils.py:188  _build_source_builder
    isaaclab/scene/interactive_scene.py:187  __init__

Local verification:

Check	Result
test_parse_env_cfg_presets.py (new)	3 passed — fails on develop
kit-less collection, no isaacsim import	58 tests in 7.3s
Isaac-Reach-Franka, -OSC, Isaac-Reorient-Cube-Shadow-Direct	6 passed
Isaac-Velocity-Flat-G1, -Spot on Newton for the first time	4 passed

Why not raise the usd-core pin

usd-core>=26.5 is the real fix, but it cannot be delivered yet: newton[importers] declares
usd-core>=25.5,<26.5, whose exclusive bound excludes the fixing release. [tool.uv] override-dependencies gets uv sync past that, but isaaclab.sh --install resolves
[project.dependencies] with pip, which has no override mechanism (see the comment at
source/isaaclab/isaaclab/cli/commands/install.py:341), so the Docker image build fails with
ResolutionImpossible.

Raising usd-core<26.5 to <27 in newton[importers] is the prerequisite. Once that lands, the
pin can move and both this workaround and the one in rendering_test_utils.py can be removed.

Notes for review

• This is a mitigation for the crash, not a fix. Users creating Newton environments outside this
  suite remain exposed until the pin can move.
• Change 3 alters what CI exercises: any task whose default is not newton_mjwarp now genuinely runs
  Newton in this suite. G1 and Spot pass; other suites are unaffected because presets defaults to
  an empty selection.
• apply_overrides still accepts a presets argument that it never reads. It no longer has callers
  in env_test_utils.py, but rendering_test_utils.py and test_hydra.py still pass it. Removing
  it is a separate cleanup across those call sites.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist

• Pre-commit hooks pass
• Changelog fragment added
• Full CI validation

[greptile-apps]

Stale USD version assertions

When the metadata and benchmark tests run after this dependency bump, they still require the old usd-core>=25.11,<26.0 constraint and installed version 25.11, causing deterministic assertion failures with the newly resolved 26.8 version.

[greptile-apps]

Obsolete single-thread USD workaround

The rendering test utility still forces PXR_WORK_THREAD_LIMIT=1 even though its TODO names usd-core 26.5 as the removal threshold. Leaving it in place slows rendering tests and prevents them from exercising normal concurrent USD parsing.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[greptile-apps]

Greptile Summary

Stabilizes Newton environment smoke testing and moves it to the kit-less path.

• Limits OpenUSD worker concurrency to avoid the parser race on the currently pinned USD release.
• Resolves requested physics presets directly in parse_env_cfg.
• Filters smoke-test tasks using the resolved configuration’s Kit requirement.
• Adds focused coverage for default and explicitly selected physics presets.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
source/isaaclab_tasks/isaaclab_tasks/utils/parse_cfg.py	Adds preset selection to environment configuration resolution while preserving default behavior for existing callers.
source/isaaclab_tasks/test/env_test_utils.py	Uses resolved presets for execution and introduces configuration-derived filtering for Kit requirements.
source/isaaclab_tasks/test/core/test_environments_newton.py	Converts the Newton smoke suite to kit-less execution and applies the OpenUSD concurrency workaround before collection.
source/isaaclab_tasks/test/core/test_parse_env_cfg_presets.py	Verifies default preset resolution, explicit Newton selection, and compatibility with other configuration overrides.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A["Collect registered task"] --> B["Load raw environment config"]
    B --> C["Resolve newton_mjwarp preset"]
    C --> D["Scan resolved config for Kit requirement"]
    D -->|Kit required| E["Exclude from kit-less suite"]
    D -->|Kit-less| F["Create environment with resolved preset"]
    F --> G["Run Newton smoke test"]

Loading

_{Reviews (2): Last reviewed commit: "Run the Newton environment smoke tests k..." | Re-trigger Greptile}

[AntoineRichard]

Some comments could be shorter, but if it works it works.

[isaaclab-review-bot]

Isaac Lab Review Bot

The preset-selection plumbing and kit-less Newton smoke-test path are coherent, but the changelog fragment incorrectly classifies the PR as test-only despite changing the public parse_env_cfg API and behavior.

• Design and architecture: The kit-less split is derived from sim_launcher.scan(...).needs_kit, keeping task selection aligned with launcher behavior rather than a separate maintained list. Applying preset selection during PresetCfg resolution also removes the previous ineffective post-resolution override path.
• API: parse_env_cfg gains a backward-compatible optional presets parameter, but this public API and behavior change is hidden by a .skip changelog fragment. Replace it with an appropriate .rst fragment describing the new preset-selection support; the current fragment also claims a file split that is not present in the patch.
• Implementation: Preset resolution occurs before the existing device and environment-count overrides, and the Isaac Sim version check now short-circuits correctly for kit-less execution. The module-scope OpenUSD worker limit is explicitly documented as a temporary mitigation with a removal condition.

Minor fixes needed. Posted 1 actionable finding inline.

Automated review; human maintainers own approval decisions.

[isaaclab-review-bot]

🟡 Warning · Api — Public API change filed as skip fragment

The fragment uses the .skip tier, which the repository rules reserve for CI/docs/test-only changes and which yields no changelog entry and no version bump. This PR also changes public source: parse_env_cfg gains a presets parameter and new preset-resolution behavior. File an .rst fragment for isaaclab_tasks with an Added/Fixed entry describing parse_env_cfg(..., presets=...). The text also mentions a file split not present in this diff.