Fix quaternion convention in locomanip SDG lerobot conversion dataset

[jaybdub]

Description

scripts/imitation_learning/locomanipulation_sdg/gr00t/convert_dataset.py
still passed scalar_first=True to Rotation.from_quat / as_quat, but
since #4437 (Jan 2026) Isaac Lab returns body poses in XYZW. The script
therefore mis-interpreted recorded poses as WXYZ, corrupting every
relative-pose computation (hands, object, fixture, goal) in the GR00T
training data and causing trained policies to produce twisted-arm
rollouts.

Type of change

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

Screenshots

Please attach before and after screenshots of the change if applicable.

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[isaaclab-review-bot]

🤖 Isaac Lab Review Bot

PR #5754: Fix quaternion convention in locomanip SDG lerobot conversion dataset

Summary

This PR fixes a quaternion convention mismatch in the locomanipulation SDG dataset conversion script. After PR #4437 changed Isaac Lab from WXYZ to XYZW convention (Jan 2026), this script was still using scalar_first=True when calling scipy's Rotation.from_quat() and as_quat(), causing all converted datasets to have corrupted pose data.

Analysis

✅ Correctness of the Fix

The fix is correct. Here's the breakdown:

Convention	Quaternion Order	scipy scalar_first
WXYZ (old Isaac Lab)	(w, x, y, z)	True
XYZW (current Isaac Lab)	(x, y, z, w)	False (default)

Since Isaac Lab now uses XYZW (scalar-last) and scipy's from_quat() / as_quat() default to scalar_first=False, removing the explicit scalar_first=True parameter correctly aligns the code with the new convention.

✅ Documentation Updates

The docstrings are properly updated:

• pose_to_transform: Updated to document "scalar-last (x, y, z, w) order" and references isaaclab.assets.ArticulationData.body_link_pose_w
• pose_from_transform: Updated to match the same convention documentation

⚠️ Minor Observations

1. No regression test included: The PR description checklist shows testing is not checked. For a bug fix affecting training data quality, a unit test verifying the conversion matches expected output would be valuable.

2. Downstream impact: Since this changes the output format for all converted datasets, any existing converted datasets would need to be regenerated. This is expected behavior for a bug fix.

Checklist Review

Item	Status
Bug fix classification	✅ Correct
Quaternion math	✅ Correct
Docstrings	✅ Updated
Pre-commit checks	⏳ Not confirmed in checklist
Tests added	❌ Not added
Changelog updated	❌ Not checked

Recommendation

The core fix is technically correct and addresses a real data corruption issue. Consider adding:

1. A simple unit test to guard against future convention regressions
2. Changelog fragment for this bug fix

---

This review was generated by the Isaac Lab Review Bot. CI checks are currently pending.

[greptile-apps]

Greptile Summary

This PR fixes a quaternion convention mismatch in scripts/imitation_learning/locomanipulation_sdg/gr00t/convert_dataset.py. Since PR #4437 (Jan 2026), Isaac Lab returns body poses with quaternions in XYZW (scalar-last) order, but the conversion script was still passing scalar_first=True to scipy's Rotation.from_quat/as_quat, causing every relative-pose computation to silently misinterpret the quaternion components and corrupt GR00T training data.

• Removes scalar_first=True from Rotation.from_quat and rotation.as_quat(), aligning the script with Isaac Lab's current XYZW convention.
• Updates docstrings for pose_to_transform and pose_from_transform to explicitly document the scalar-last convention and references body_link_pose_w for traceability.

Confidence Score: 5/5

The change is a two-line fix with no side-effects — it removes an incorrect flag from two scipy calls and updates matching docstrings. The rest of the codebase (rollout_policy.py) already operates with XYZW quaternions, confirming the direction is correct.

Both the conversion direction (from_quat) and output direction (as_quat) are updated consistently. The rollout script independently establishes that the environment emits XYZW and defaults to that format, corroborating the fix. No other files in the scripts/imitation_learning tree use scalar_first.

No files require special attention.

Important Files Changed

Filename	Overview
scripts/imitation_learning/locomanipulation_sdg/gr00t/convert_dataset.py	Removes scalar_first=True from Rotation.from_quat/as_quat in pose_to_transform and pose_from_transform, correctly aligning quaternion parsing with Isaac Lab's XYZW convention; docstrings updated accordingly.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["HDF5 Dataset\n(Isaac Lab XYZW poses)"] --> B["pose_to_transform()\nRotation.from_quat(q)\ndefault: scalar_first=False = XYZW ✅"]
    B --> C["RigidTransform\nbatch Rotation"]
    C --> D["compute_relative_pose()\nbase_pose.inv() * target"]
    D --> E["pose_from_transform()\nrotation.as_quat()\ndefault: scalar_first=False = XYZW ✅"]
    E --> F["7D pose array\n[x,y,z, x,y,z,w]"]
    F --> G["Parquet / GR00T\nTraining Dataset"]

Loading

_{Reviews (1): Last reviewed commit: "fix quaternion scalar_first=True bug in ..." | Re-trigger Greptile}