Fix LEAPP projected gravity export

[lgulich]

Summary

The current LEAPP export exports projected_gravity as an input instead of root_quat_w.
The reason is that with the change to use warp for observation terms LEAPP can no longer automatically trace projected gravity back to root_quat_w. For new we fix this manually, once we have warp support in LEAPP this will no longer be needed.

Changes

• Export mdp.projected_gravity through root_quat_w for LEAPP graphs.
• Keep the policy observation as projected gravity by computing it inside the exported graph.
• Add regression coverage and a changelog fragment.

Test Plan

• uv run --with pytest --with leapp python -m pytest source/isaaclab_rl/test/export/test_leapp_proxy.py -q
• uv run --with ruff --with leapp ruff check source/isaaclab/isaaclab/utils/leapp/export_annotator.py source/isaaclab/isaaclab/utils/leapp/proxy.py source/isaaclab_rl/test/export/test_leapp_proxy.py
• ./isaaclab.sh -f
• Exported Isaac-Velocity-Flat-G1-v0 checkpoint model_1499.pt and verified YAML uses state/body/rotation for robot_root_quat_w.

[isaaclab-review-bot]

Code Review Summary

This PR correctly fixes the LEAPP export of projected_gravity by exposing the root quaternion orientation (root_quat_w) at the graph boundary instead of the pre-computed projected gravity vector. This is the right architectural decision since deployment backends (robot hardware/simulation) typically provide body orientation from IMUs rather than pre-computed projected gravity.

✅ Strengths

1. Sound Architecture: Exposing root_quat_w as a BODY_ROTATION input instead of projected_gravity as VECTOR3D aligns with what deployment backends actually provide. The projection computation happens inside the exported graph where it belongs.

2. Correct Quaternion Math: The inline quaternion inverse rotation implementation correctly matches the existing math_utils.quat_apply_inverse formula and properly handles Isaac Sim's (x, y, z, w) scalar-last quaternion convention.

3. Good Test Coverage: The new tests verify both the direct proxy path (preserving VECTOR3D semantics) and the observation export path (using root_quat_w with BODY_ROTATION semantics), with mathematical correctness validated against the reference implementation.

4. Proper Changelog Entry: Follows the project's changelog conventions.

📝 Follow-up Review (d30f2b7)

The new commits since the initial review (30d66c0→d30f2b7) contain only housekeeping changes:

• Documentation updates to use the new task naming convention (Isaac-Reach-UR10 instead of Isaac-Reach-UR10-v0)
• Test file rename and task name updates to match the codebase-wide rename
• Changelog .skip marker added

No changes to the core LEAPP projected gravity fix logic. The suggestions from the initial review remain optional enhancements and do not block merge.

📋 Overall Assessment

This is a well-implemented fix that addresses a real deployment concern. The approach is architecturally sound, the math is correct, and the test coverage validates the key behaviors. Ready to merge. ✅

---

Update (e290070): New commits are a merge of develop into the feature branch (cloner refactor #5770, OvPhysX runtime fix #5995, CI changelog bump). No changes to the LEAPP projected gravity fix. Previous assessment stands — ready to merge.

---

Update (2bd4a4b): New commit fixes hidden_state → hidden_states attribute name in two places in export.py (lines 202 and 332). Straightforward correctness fix aligning with the actual API. No new issues introduced. Previous assessment unchanged — ready to merge. ✅

---

Update (ae5481a): Merge of develop into the feature branch — brings in docs command updates (#6023) and PPISP optional-dependency refactor (#5986). No changes to LEAPP export logic or tests. No new issues. ✅

---

Update (6f18d1c): This commit reverts the previous hidden_states back to hidden_state (singular) in export.py (lines 202 and 332). The commit message explains the rationale well: rsl_rl 5.0.1 (which Isaac Lab pins) uses hidden_state (singular), while the plural form (hidden_states) only exists in older/vendored forks like AGILE's 2.3.3.

This is a good catch — the earlier commit (2bd4a4b) was incorrect for the pinned rsl_rl version. The revert restores the correct attribute name. Net effect: the export.py file is unchanged from its state before the 2bd4a4b commit, which is correct for rsl_rl 5.0.1.

No new issues. Core LEAPP projected gravity fix remains intact. ✅

[frlai]

can you move the tests to: IsaacLab/source/isaaclab_rl/test/export. that's where all the leapp export tests are.

[greptile-apps]

Greptile Summary

This PR fixes the LEAPP graph export for the projected_gravity observation by surfacing root_quat_w (body orientation) as the graph input instead of an already-projected vector, then re-deriving the projected gravity inside the exported graph.

• ExportPatcher._wrap_projected_gravity is added: when the observation manager is patched, any projected_gravity term is replaced by a wrapper that reads root_quat_w through the LEAPP proxy (annotating it as a BODY_ROTATION input) and applies the inverse quaternion rotation to [0, 0, -1] to reconstruct the body-frame gravity vector.
• Two regression tests are added in source/isaaclab_rl/test/export/test_leapp_proxy.py: one verifying the direct _DataProxy path retains its VECTOR3D semantic, and one verifying the export-patched path emits the correct BODY_ROTATION semantic and produces numerically correct gravity projections.

Confidence Score: 4/5

Safe to merge; the quaternion convention and rotation formula are correct, and regression tests validate both the annotated and unannotated paths.

The core math (xyzw convention, Rodrigues inverse-rotation formula) matches the existing quat_apply_inverse helper exactly, the asset-name fallback to robot handles the default-config case correctly, and the tests assert numerical correctness against the reference implementation. The only things worth noting are a minor duplication of the rotation formula and a stale file path in the PR description test-plan commands that could confuse contributors running the tests manually.

The new test file at source/isaaclab_rl/test/export/test_leapp_proxy.py — verify the CI configuration picks it up from that location rather than the path quoted in the PR description.

Important Files Changed

Filename	Overview
source/isaaclab/isaaclab/utils/leapp/export_annotator.py	Adds _wrap_projected_gravity which exports root orientation as root_quat_w and re-derives the projected gravity in-graph using a hand-rolled quat_apply_inverse; formula and quaternion convention (x,y,z,w) are correct.
source/isaaclab_rl/test/export/test_leapp_proxy.py	New regression test validating both the direct _DataProxy.projected_gravity_b path (VECTOR3D semantic) and the export-patched path (BODY_ROTATION semantic via root_quat_w); test logic is correct but file is at a different path than the PR description's test-plan commands.
source/isaaclab/changelog.d/lgulich-leapp-root-quat-gravity.rst	Changelog fragment documenting the fix; content is accurate.
source/isaaclab_rl/changelog.d/lgulich-leapp-root-quat-gravity.skip	Empty skip file indicating no changelog entry is needed for the isaaclab_rl package.

Sequence Diagram

sequenceDiagram
    participant ObsManager as Observation Manager
    participant Patcher as ExportPatcher
    participant Wrapper as _wrap_projected_gravity
    participant ProxyEnv as _EnvProxy / _DataProxy
    participant LEAPP as LEAPP annotate

    ObsManager->>Patcher: _patch_observation_manager(obs_manager, proxy_env)
    Patcher->>Patcher: "detect func_name == projected_gravity"
    Patcher->>Wrapper: _wrap_projected_gravity(original_func, proxy_env)
    Patcher-->>ObsManager: "replace term_cfg.func, set noise=None"

    Note over ObsManager: At policy rollout / export trace
    ObsManager->>Wrapper: term_cfg.func(env, [asset_cfg])
    Wrapper->>ProxyEnv: proxy_env.scene[asset_name].data.root_quat_w.torch
    ProxyEnv->>LEAPP: annotate.input_tensors(task_name, BODY_ROTATION semantics)
    LEAPP-->>ProxyEnv: annotated root_quat_w tensor
    ProxyEnv-->>Wrapper: root_quat_w (Nx4, x,y,z,w)
    Wrapper->>Wrapper: "gravity_w = [0,0,-1] broadcast to (N,3)"
    Wrapper->>Wrapper: "projected_gravity_b = quat_apply_inverse(root_quat_w, gravity_w)"
    Wrapper-->>ObsManager: projected_gravity_b (Nx3)

Loading

_{Reviews (1): Last reviewed commit: "Merge branch 'develop' into lgulich/leap..." | Re-trigger Greptile}