[Backport release/3.0.0] Fix Newton body-offset Jacobian and streamline core offset math (#8203) by kellyguo11 · Pull Request #8318 · isaac-sim/IsaacLab

kellyguo11 · 2026-10-06T04:29:40Z

Description

Backports #8203 onto release/3.0.0.

This fixes the Newton task-space body-offset Jacobian by shifting its translational rows with the world-frame lever arm before rotating the Jacobian into root axes. It also carries the equivalent core task-space optimization that replaces skew-matrix multiplication with batched cross products.

This is an exact, conflict-free cherry-pick of the merged source change. The Newton regression coverage and both package changelog fragments are included unchanged.

Type of change

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

Release backport

Not applicable: this PR targets release/3.0.0 directly.

Screenshots

Not applicable.

Validation

Exact-patch validation passed: the source and backport patch IDs both resolve to 1a90740c29ae269d3409b448dd321d67ba4d2a55.
Release-aware changelog validation passed against the current upstream release/3.0.0 tip.
All repository pre-commit hooks passed over the backport range.
git diff --check passed.
Test-audit review confirmed that the consolidated Newton controller test protects the Newton-specific body-offset frame convention and would catch the original world/root-frame mismatch without adding a production test seam.
Fix Newton body-offset Jacobian and streamline core offset math #8203 reported 8 Newton controller tests and 5 core DiffIK tests passing. The focused test was not repeated locally because the project lockfile supports Linux and Windows, not this macOS ARM64 host.

Checklist

I have read and understood the contribution guidelines.
I have run the applicable formatting and lint checks.
Documentation changes are not required beyond the included changelog entries.
My changes generate no new warnings.
I have validated the changed behavior with focused coverage and applied the test-audit criteria.
I have included the required isaaclab and isaaclab_newton changelog fragments.
The original contributor is already listed in CONTRIBUTORS.md.

Original PR: #8203
Cherry-picked merge commit: e67c5cb4cfbc8f40de8bf60559048707ea229f58
Backport commit: e3fc1ed20

greptile-apps · 2026-10-06T04:32:01Z

Confidence Score: 4/5

[Medium risk] Refactors task-space Jacobian math for offset frames.

The PR appears safe to merge; the test assertion is a non-blocking maintainability issue.

Findings

Unrelated assertion makes test fragile ▶

Summary

This backport corrects the Newton body-offset Jacobian by applying the lever-arm shift in world coordinates before rotating into root axes. It also replaces skew-matrix products with batched cross products in the core task-space actions and adds regression coverage and changelog fragments.

_{Reviews (1) · Last reviewed commit: "Fix Newton body-offset Jacobian and stre..."}

greptile-apps · 2026-10-06T04:32:04Z

+    # The body-local x offset is root-local y; a root-z rotation gives linear velocity -x.
+    expected = torch.tensor([[-1.0, 0.0, 0.0, 0.0, 0.0, 1.0]]).repeat(_NUM_ENVS, 1)
+    torch.testing.assert_close(jacobian[:, :, 0], expected, atol=1e-6, rtol=0.0)
+    assert not hasattr(math_utils, "velocity_at_point")


Unrelated assertion makes test fragile

This assertion checks that a math helper does not exist, rather than checking the Jacobian or controller result. If velocity_at_point is added later, this test will fail even though the behavior it tests remains correct. The Jacobian and controller results are already checked separately; remove this assertion.

Suggested change

assert not hasattr(math_utils, "velocity_at_point")

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!

isaaclab-review-bot

Isaac Lab Review Bot

Needs attention: Test quality (1). Other reviewed categories look good.

PR description: The “reduced overhead” claim lacks measurements; cite #8203 with baseline/candidate revisions, command, hardware, backend, environment count, warmup, and metric, or remove the performance wording.
Design and architecture: Architecture looks good.
Scope and over-engineering: No over-engineering issues.
API: API changes look good.
Compatibility and deprecation: No breaking changes identified.
Implementation: Implementation looks good.
Style consistency: Style consistency looks good.
Test quality: The Jacobian regression covers the Newton-specific frame mismatch, but the unrelated velocity_at_point symbol-absence assertion is implementation inventory and must be removed.

Minor fixes needed. Posted 1 inline finding.

Automated review; human maintainers own approval decisions.

isaaclab-review-bot · 2026-10-06T04:33:53Z

+    # The body-local x offset is root-local y; a root-z rotation gives linear velocity -x.
+    expected = torch.tensor([[-1.0, 0.0, 0.0, 0.0, 0.0, 1.0]]).repeat(_NUM_ENVS, 1)
+    torch.testing.assert_close(jacobian[:, :, 0], expected, atol=1e-6, rtol=0.0)
+    assert not hasattr(math_utils, "velocity_at_point")


🟡 Warning · Test Quality — Unrelated symbol-absence assertion in test

assert not hasattr(math_utils, "velocity_at_point") guards no Jacobian contract and fails if any unrelated helper gains that name. Delete it, correcting the source change on main so this backport stays an exact cherry-pick.

kellyguo11 · 2026-10-06T04:34:37Z

run-ci

# Description Enable the root `ov` extra in the release base Docker image so its locked environment includes `ovphysx`, `ovrtx`, and `ovstage`. This aligns `release/3.0.0` with the base-image dependency set on `develop` and fixes OVRTX-backed task configurations failing during standard Docker package tests with `ModuleNotFoundError: No module named 'ovrtx'`.\n\nThis is a focused release-branch workflow fix for the CI failure exposed by #8318. It does not change dependency pins or the lockfile.\n\n## Type of change\n\n- Bug fix (non-breaking Docker dependency fix)\n\n## Release backport\n\nNot applicable; this PR targets `release/3.0.0` directly.\n\n## Screenshots\n\nNot applicable.\n\n## Validation\n\n- `uv run --extra test python -m pytest docker/test/test_dockerfile_nonroot.py` — 24 passed\n- `SKIP=check-changelog-fragments uv run isaaclab -f` — all applicable formatting and lint hooks passed\n- `git diff --check`\n\nThe full formatting command’s changelog hook compares the release snapshot against `develop` and reports the release branch’s already-compiled fragments. No `source/<package>` files are changed, so no package changelog fragment applies.\n\n## Checklist\n\n- [x] I have read and understood the contribution guidelines\n- [x] I have run the applicable formatting and lint checks\n- [x] Documentation changes are not required for this Docker dependency fix\n- [x] My changes generate no new warnings\n- [x] I have validated the closest Docker contract tests\n- [x] No source package changed, so no package changelog fragment is required\n- [x] My name already exists in `CONTRIBUTORS.md` --------- Co-authored-by: Octi Zhang <zhengyuz@nvidia.com>

AntoineRichard · 2026-10-06T08:30:37Z

run-ci

…c-sim#8203) # Description A task-space action with `body_offset` needs a Jacobian at the offset point. The core DiffIK and OSC actions received the corrected frame convention in isaac-sim#7989, but the Newton model-free action retained its own formula: it used a body-local lever arm against a root-frame Jacobian and rotated the angular rows by the offset orientation. With a rotated body or root, `J @ qdot` therefore disagreed with the offset-frame twist. This PR fixes the Newton action by rotating the offset into world axes, applying `Jv_point = Jv_origin + Jw × r`, and then rotating the Jacobian into root axes. The angular rows stay unchanged. Core DiffIK and OSC now use the same direct `torch.linalg.cross` operation in root axes instead of constructing a skew matrix and multiplying it by the angular Jacobian. The two forms are mathematically equivalent; the direct cross product avoids an extra matrix allocation and multiplication. No new math wrapper is needed. The Newton task-space actions use Newton's model-free controllers, which require a caller-supplied tool-point pose and Jacobian. Newton's separate model-based controllers and IK objectives support tool-site/link offsets, but are not the controller API used here. The Newton action continues to accept both Newton and PhysX asset data through the public link-origin Jacobian contract. On the official Franka validation from the original fix (rotated fixed root and a hand offset), `J @ qdot` matched the simulator's offset-frame twist: Newton linear error fell from about 0.27 m/s to 2e-6 m/s and angular error from about 0.95 rad/s to 1e-7 rad/s. PhysX errors were 1e-4 to 4e-4, matching the core term's floor. Only configurations with `body_offset` are affected; no shipped task currently uses these Newton action terms. Fixes isaac-sim#8202 ## Type of change - Bug fix and task-space offset performance improvement ## Release backport - [x] Backport this pull request to the active release branch after it merges into `develop` ## Validation - `uv run --frozen --extra test python -m pytest source/isaaclab/test/envs/test_diffik_jacobian_aliasing.py -q` — 5 passed. - `uv run --frozen --extra test python -m pytest source/isaaclab_newton/test/controllers/test_newton_controllers.py -q` — 8 passed; the existing DiffIK case now checks the offset Jacobian. - `uv run --frozen isaaclab -f` — passed with the local changelog base set to this PR's upstream base commit. - 4,096-env, 7-joint Torch GPU microbenchmarks on RTX 5090 (median of 8 paired rounds of 100 calls): Newton world-frame point shift, direct cross 149.6 us/call versus skew matrix 213.5 us/call; core root-frame point shift, direct cross 107.0 us/call versus skew matrix 181.0 us/call. These measure Jacobian computation, not a full simulation step. ## Checklist - [x] I have read and understood the contribution guidelines. - [x] I have run the pre-commit checks. - [x] I have added changelog fragments for both touched packages. - [x] The focused regression tests pass. - [ ] I have made corresponding changes to the documentation. --------- Co-authored-by: Octi Zhang <zhengyuz@nvidia.com> (cherry picked from commit e67c5cb)

kellyguo11 · 2026-10-06T14:56:06Z

run-ci

kellyguo11 requested a review from a team October 6, 2026 04:29

kellyguo11 added the isaac-lab Related to Isaac Lab team label Oct 6, 2026

kellyguo11 requested review from AntoineRichard, StafaH, fatimaanes, hujc7, ooctipus and pbarejko as code owners October 6, 2026 04:29

greptile-apps Bot reviewed Oct 6, 2026

View reviewed changes

isaaclab-review-bot Bot reviewed Oct 6, 2026

View reviewed changes

isaaclab-bot Bot added ci:run-docker Trigger the on-demand Docker and GPU CI workflow and removed ci:run-docker Trigger the on-demand Docker and GPU CI workflow labels Oct 6, 2026

kellyguo11 mentioned this pull request Oct 6, 2026

[Release 3.0.0] Enable OV runtime in base Docker image #8321

Merged

AntoineRichard approved these changes Oct 6, 2026

View reviewed changes

isaaclab-bot Bot added ci:run-docker Trigger the on-demand Docker and GPU CI workflow and removed ci:run-docker Trigger the on-demand Docker and GPU CI workflow labels Oct 6, 2026

AntoineRichard force-pushed the codex/backport-8203-release-3.0.0 branch from e3fc1ed to a62bc47 Compare October 6, 2026 08:53

Merge branch 'release/3.0.0' into codex/backport-8203-release-3.0.0

b41e40c

isaaclab-bot Bot added ci:run-docker Trigger the on-demand Docker and GPU CI workflow and removed ci:run-docker Trigger the on-demand Docker and GPU CI workflow labels Oct 6, 2026

kellyguo11 merged commit ed73bb9 into isaac-sim:release/3.0.0 Oct 6, 2026
57 of 58 checks passed

Navigation Menu

Repository navigation

[Backport release/3.0.0] Fix Newton body-offset Jacobian and streamline core offset math (#8203) - #8318

[Backport release/3.0.0] Fix Newton body-offset Jacobian and streamline core offset math (#8203)#8318
kellyguo11 merged 2 commits into
isaac-sim:release/3.0.0from
kellyguo11:codex/backport-8203-release-3.0.0

kellyguo11 commented Oct 6, 2026

Uh oh!

greptile-apps Bot commented Oct 6, 2026 •

edited

Loading

Uh oh!

greptile-apps Bot Oct 6, 2026

Uh oh!

isaaclab-review-bot Bot left a comment

Uh oh!

isaaclab-review-bot Bot Oct 6, 2026

Uh oh!

kellyguo11 commented Oct 6, 2026

Uh oh!

AntoineRichard commented Oct 6, 2026

Uh oh!

kellyguo11 commented Oct 6, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

kellyguo11 commented Oct 6, 2026

Description

Type of change

Release backport

Screenshots

Validation

Checklist

Uh oh!

greptile-apps Bot commented Oct 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 4/5

Findings

Summary

Uh oh!

greptile-apps Bot Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

isaaclab-review-bot Bot left a comment

Choose a reason for hiding this comment

Isaac Lab Review Bot

Uh oh!

isaaclab-review-bot Bot Oct 6, 2026

Choose a reason for hiding this comment

Uh oh!

kellyguo11 commented Oct 6, 2026

Uh oh!

AntoineRichard commented Oct 6, 2026

Uh oh!

kellyguo11 commented Oct 6, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

greptile-apps Bot commented Oct 6, 2026 •

edited

Loading