Pin Newton to v1.4.0 and override the Isaac Sim MuJoCo pins

[hujc7]

Summary

• Advances the Newton pin from a pre-1.4 development commit (c7ae7c7648c) to the v1.4.0 release, which contains the ignore_paths custom-frequency USD traversal fix (newton#3406, membership in the release verified) plus the 1.4 stabilization series — 50 commits newer than the previously targeted dev SHA.
• Newton 1.4 requires the MuJoCo 3.10 stack while isaacsim-core 6.0.0.1 pins mujoco-warp==3.8.0.3 exactly, so mujoco-warp>=3.10.0.2,<3.11 and mujoco~=3.10.0 join the [tool.uv] override-dependencies (the same mechanism already used to out-rank isaacsim's torch and newton pins; the 3.10.0.2 floor is what v1.4.0 declares). The version table and its single-source test carry the new entries.
• Extracted from [Task Clean-up][Newton] Dexterous Part 1/8: Fix cloner rows, cubric fallback, and visualizer teardown #6411 so the approved runtime fixes there stay decoupled from the dependency advance: [Task Clean-up][Newton] Dexterous Part 1/8: Fix cloner rows, cubric fallback, and visualizer teardown #6411's Multi-GPU CI failure bisects to this pin, not to its code.

Problem: mesh-mesh contact regression in the MuJoCo pipeline (upstream)

The Newton range this pin crosses includes newton#2980 (MuJoCo 3.10 margin/gap semantics), which introduces a regression: mesh-mesh shape pairs in the MuJoCo contacts pipeline lose collision response entirely — bodies tunnel through each other with zero contact forces.

Bisect (via test_horizontal_collision_detects_contact[mesh_capsule-mujoco_contacts]):

Newton ref	Result
c7ae7c7648c (current develop pin)	pass
346121aaaca = newton#2980	fail (first bad)
9af5a9f4181 (1.5.0.dev0)	fail
v1.4.0	fail

• No older pin avoids it: the ignore_paths fix landed 31 commits after the regression, so every Newton commit containing the fix also contains the regression.
• legacy_margin_gap=True at USD import does not restore the behavior (the failure is contact-pair loss, not margin-value translation), and neither does raising the builder default gap.
• Scope: primitive pairs and mesh-vs-primitive pairs are unaffected (the dexterous task families train to full success rates under this pin).
• Tracked upstream as newton#3559; the affected test case is marked xfail referencing that issue and should be re-enabled when the upstream fix lands.

Stacking

• Independent of the dexterous task-cleanup series (based on develop). If [Task Clean-up][Newton] Dexterous Part 1/8: Fix cloner rows, cubric fallback, and visualizer teardown #6411 merges first, this PR picks up the removal of its now-redundant ignore_paths cloner workaround on rebase.

[greptile-apps]

Greptile Summary

This PR advances the Newton dependency from a pre-1.4 dev commit to the v1.4.0 release tag, bringing in the ignore_paths USD traversal fix, and adds mujoco-warp>=3.10.0.2,<3.11 and mujoco~=3.10.0 as uv override-dependencies to satisfy Newton 1.4's MuJoCo 3.10 requirement over Isaac Sim's pinned 3.8 stack.

• pyproject.toml: Newton pin changed to v1.4.0; two new version-table entries (mujoco_warp, mujoco) and matching override-dependencies entries added using the same override mechanism already in use for torch and newton.
• test_uv_run_pyproject.py: Two new assertions verify that the mujoco-warp/mujoco override strings in [tool.uv] match the version table, keeping the single-source-of-truth invariant intact.
• test_contact_sensor.py: Marks the mesh_capsule + mujoco_contacts combination as xfail with a reference to upstream newton#3559, reflecting a known mesh-mesh contact regression introduced by the MuJoCo 3.10 margin/gap semantics change.

Confidence Score: 4/5

Safe to merge as a dependency advance; the main question to resolve is whether the xfail guard in test_contact_sensor.py covers the full extent of the broken mesh-mesh combinations.

The pyproject.toml dependency wiring and version-table test are correct. The concern is in test_contact_sensor.py: the xfail only guards MESH_CAPSULE, but the upstream regression is described as affecting all mesh-mesh pairs, and STABLE_SHAPES includes MESH_SPHERE, MESH_BOX, and MESH_CYLINDER as well. If those parametrised cases also fail under MuJoCo 3.10 contacts, CI would see hard failures on them rather than expected failures.

source/isaaclab_newton/test/sensors/test_contact_sensor.py — the xfail condition should be verified against all mesh shape types, not just MESH_CAPSULE.

Important Files Changed

Filename	Overview
pyproject.toml	Advances Newton pin to v1.4.0 release tag and adds mujoco-warp/mujoco 3.10 overrides to both the version table and uv override-dependencies; stale "git commit" comment is the only minor issue.
source/isaaclab/test/cli/test_uv_run_pyproject.py	Adds two assertions to the single-source version test that verify the new mujoco-warp and mujoco overrides match the version table; logic is correct and consistent with the existing assertion pattern.
source/isaaclab_newton/test/sensors/test_contact_sensor.py	Adds an xfail guard for mesh_capsule + mujoco_contacts, but the regression (described as affecting all mesh-mesh pairs) may leave MESH_SPHERE, MESH_BOX, and MESH_CYLINDER uncovered; docstring placement is also incorrect.
source/isaaclab_newton/changelog.d/newton-pin-v140.minor.rst	New changelog fragment documenting the Newton v1.4.0 pin and the ignore_paths fix; content is accurate and complete.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["isaacsim-core 6.0.0.1\n(pins mujoco-warp==3.8.0.3, mujoco==3.8.x, newton==1.2.0)"]
    B["[tool.uv] override-dependencies\n(Isaac Lab wins over every requester)"]
    C["newton[sim] @ git+...@v1.4.0"]
    D["mujoco-warp >=3.10.0.2,<3.11"]
    E["mujoco ~=3.10.0"]
    F["torch==2.11.0 stack"]
    G["Resolved environment\n(Newton 1.4 + MuJoCo 3.10)"]
    A -->|"exact pins blocked by"| B
    B --> C
    B --> D
    B --> E
    B --> F
    C --> G
    D --> G
    E --> G
    F --> G

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["isaacsim-core 6.0.0.1\n(pins mujoco-warp==3.8.0.3, mujoco==3.8.x, newton==1.2.0)"]
    B["[tool.uv] override-dependencies\n(Isaac Lab wins over every requester)"]
    C["newton[sim] @ git+...@v1.4.0"]
    D["mujoco-warp >=3.10.0.2,<3.11"]
    E["mujoco ~=3.10.0"]
    F["torch==2.11.0 stack"]
    G["Resolved environment\n(Newton 1.4 + MuJoCo 3.10)"]
    A -->|"exact pins blocked by"| B
    B --> C
    B --> D
    B --> E
    B --> F
    C --> G
    D --> G
    E --> G
    F --> G

Loading

_{Reviews (1): Last reviewed commit: "Reference the upstream Newton issue from..." | Re-trigger Greptile}

[greptile-apps]

xfail scope may be too narrow for other mesh types

The PR description characterises the regression as "mesh-mesh shape pairs in the MuJoCo contacts pipeline lose collision response entirely," yet the guard only covers MESH_CAPSULE. STABLE_SHAPES also includes MESH_SPHERE, MESH_BOX, and MESH_CYLINDER. Because test_horizontal_collision_detects_contact runs the same shape_type for both objects, those combinations are also mesh-mesh collisions and are likely affected. If any of them fail under the new MuJoCo 3.10 stack they will produce an unguarded hard failure in CI rather than an expected-failure, which would block merges unrelated to this pin. Consider broadening the condition to use_mujoco_contacts and is_mesh_shape(shape_type) (the helper is already imported) and re-evaluating once the upstream fix in newton#3559 lands.

[greptile-apps]

The function docstring is placed after the early-exit block, making it an unreachable string expression rather than a proper __doc__ attribute. Python only recognises a triple-quoted string as a docstring when it is the first statement in the function body. Moving the xfail check below the docstring preserves documentation tooling support and follows the standard pytest conditional-xfail pattern.

Suggested change

	def test_horizontal_collision_detects_contact(device: str, use_mujoco_contacts: bool, shape_type: ShapeType):
	if use_mujoco_contacts and shape_type == ShapeType.MESH_CAPSULE:
	pytest.xfail(
	"Newton >= 1.4 (mujoco-warp 3.10 margin/gap semantics) loses mesh-mesh collision response in the"
	" MuJoCo contacts pipeline: objects tunnel with zero contact forces."
	" Tracked upstream: https://github.com/newton-physics/newton/issues/3559"
	)
	"""Test horizontal collision detection with varied velocities and separations.
	def test_horizontal_collision_detects_contact(device: str, use_mujoco_contacts: bool, shape_type: ShapeType):
	"""Test horizontal collision detection with varied velocities and separations.

[greptile-apps]

The comment still reads "Newton git commit" but the pin is now the v1.4.0 release tag, not a bare commit SHA.

Suggested change

	# Newton git commit + matching warp-lang prerelease (upgraded together).
	newton = "c7ae7c7648cd0717df39e5c94b95d5a02c997320"
	newton = "v1.4.0"
	# Newton release tag + matching warp-lang prerelease (upgraded together).
	newton = "v1.4.0"

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!

[AntoineRichard]

Closing this as we're going to merge this one instead: #6586