Fix Newton external wrench frame

[AntoineRichard]

Description

Newton's external-wrench writers passed the body-frame output of the wrench composer directly to Newton's world-frame State.body_f buffer. This caused forces and torques applied to rotated bodies to follow world axes instead of the requested body axes.

This change:

• rotates composed force and torque into world coordinates inside the existing Newton packing kernels;
• preserves the center-of-mass reference point by using pose rotation only;
• covers both reorder-free and body-ordered articulation paths, rigid objects, and rigid-object collections;
• adds no buffer allocation or additional kernel launch.

No new dependencies are required.

Type of change

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

Testing

Tests ran in a dedicated virtual environment created from the rebased branch's uv.lock.

• 379 passed: WrenchComposer plus direct and articulation-ordered Newton packing regressions on CPU/CUDA.
• 10 passed: rigid-object external-force integration selection.
• 18 passed: rigid-object-collection external-force integration selection.
• 20 passed: articulation external-force integration selection.
• 4 passed: focused articulation force-at-position parameterizations across CPU/CUDA and one/two environments.
• ./isaaclab.sh -f passed all hooks before commit and again before push.

The corrected wrench direction made the pre-existing 100 N articulation stress load overflow Newton's constraint capacity in three of four parameterizations. A one-line 100 N to 50 N test-load reduction keeps all 100 steps and the original angular-velocity assertion while avoiding unrelated solver divergence. Direct packing tests retain exact frame and body-order assertions.

Screenshots

Not applicable.

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh -f
• I have made corresponding documentation changes
• My changes generate no new warnings
• I have added tests that prove the fix is effective
• I have added a changelog fragment under source/isaaclab_newton/changelog.d/
• My name already exists in CONTRIBUTORS.md

[greptile-apps]

Greptile Summary

The PR corrects Newton external-wrench packing by rotating body-frame forces and torques into world coordinates using each link pose quaternion. The implementation covers shared, articulation-ordered, rigid-object, and rigid-object-collection paths and adds focused CPU Warp regression tests.

• Passes link poses into the existing wrench-packing kernels without another allocation or kernel launch.
• Preserves COM-referenced torque by using only pose rotation, not translation.
• Handles nonidentity articulation body ordering by rotating in public order before scattering into backend order.
• Adds a changelog fragment and tests for rotation, translation invariance, and body reordering.

Confidence Score: 5/5

The PR appears safe to merge with no concrete blocking or non-blocking defects identified.

The updated kernels consistently invert the wrench composer's world-to-body conversion, use correctly ordered link poses, preserve COM-referenced torque, and are covered by focused ordering and frame regressions.

Important Files Changed

Filename	Overview
source/isaaclab_newton/isaaclab_newton/assets/kernels.py	The shared packing kernel now performs the required body-to-world quaternion rotation while preserving the COM reference point.
source/isaaclab_newton/isaaclab_newton/assets/articulation/kernels.py	The ordered articulation kernel rotates public-order wrenches with matching public-order poses before scattering them into backend order.
source/isaaclab_newton/isaaclab_newton/assets/articulation/articulation.py	Both articulation packing paths now supply the link-pose buffer expected by their updated kernels.
source/isaaclab_newton/isaaclab_newton/assets/rigid_object/rigid_object.py	Rigid-object wrench packing now receives the current link pose needed for body-to-world conversion.
source/isaaclab_newton/isaaclab_newton/assets/rigid_object_collection/rigid_object_collection.py	Collection wrench packing now supplies per-object link poses to the shared conversion kernel.
source/isaaclab_newton/test/assets/test_wrench_kernels.py	Regression tests cover rotation direction, translation invariance, and nonidentity articulation body ordering.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Composed body-frame force and torque] --> B{Articulation body ordering}
    B -->|Identity or rigid asset| C[Read public-order link rotation]
    B -->|Nonidentity| D[Read public-order link rotation and body map]
    C --> E[Rotate force and torque into world frame]
    D --> E
    E --> F{Backend ordering required?}
    F -->|No| G[Write world-frame Newton wrench]
    F -->|Yes| H[Scatter to backend body index]
    H --> G

Loading

_{Reviews (1): Last reviewed commit: "Correct Newton torque unit notation" | Re-trigger Greptile}

[isaaclab-review-bot]

Isaac Lab Review Bot

The Newton external-wrench fix consistently rotates composed body-frame forces and torques into world coordinates at the packing boundary, with all changed kernel launch sites updated to pass body poses.

• Design and architecture: Performing the rotation in the Newton packing kernels preserves the wrench composer's body-frame contract and avoids additional buffers or kernel launches. The reorder-free and ordered articulation paths, rigid objects, and rigid-object collections follow the same conversion design.
• API: No public API surface is changed. The internal kernel signatures and supplied callers were updated together, while launch dimensions, masks, spatial-vector layout, and public-to-backend body ordering remain consistent. The package changelog fragment is correctly scoped and formatted.
• Implementation: The kernels use only the link-pose quaternion, correctly avoiding a translation-induced moment for COM-referenced wrenches. Public-order poses are indexed before the ordered path scatters into backend order, and regression tests cover rotation, nonzero translation, and nonidentity ordering. Asset-level integration remains a CI validation point because it was unavailable locally.

No blocking issues. No inline issue met the actionable-evidence threshold; the assessment above records the review feedback.

Automated review; human maintainers own approval decisions.

[mmichelis]

Could we centralize the body-to-world conversion in the existing update_wrench_with_force_and_torque helper and reuse it from both kernels? The helper currently only packs two vectors into wp.spatial_vector, while the ordered and unordered paths duplicate the quaternion rotations. I suggest renaming it to something explicit such as pack_body_wrench_to_world(force_b, torque_b, body_rot_w) and having it rotate both components before packing. This would keep the frame contract in one place and ensure both body-ordering paths use identical conversion math.

[StafaH]

These proxy arrays should work without .warp right?

[AntoineRichard]

They should, but I may as well use the right one explicitely!