Support OVPhysX in randomize_rigid_body_collider_offsets MDP event

[AntoineRichard]

Description

Stacked on #7563. This branch needs the REST_OFFSET / CONTACT_OFFSET / RIGID_BODY_*_OFFSET aliases and the CPU-only routing for those tensor types that #7563 adds. Review/merge that first; this diff collapses to a single commit once it lands, at which point the base should be retargeted to develop.

randomize_rigid_body_collider_offsets had no OVPhysX implementation. Its backend dispatch tested only for "newton" and fell through a bare else to _RandomizeRigidBodyColliderOffsetsPhysx, which calls get_rest_offsets / get_contact_offsets / set_rest_offsets / set_contact_offsets on asset.root_view. On OVPhysX that view is an OvPhysxView, which defines none of them, so adding the term to a config raised AttributeError during OvPhysxManager.reset() (via PHYSICS_READY) and the environment never built. The error surfaced from inside the PhysX implementation, so it did not name the unsupported backend either.

Found by the OVQA agent harness on Isaac-Cartpole (release/3.0.0, ovphysx 0.5.11). No shipped task uses the term today, so this is latent rather than a live regression.

Fix

• New _RandomizeRigidBodyColliderOffsetsOvPhysx, mirroring the PhysX variant. OVPhysX runs the PhysX solver, so rest/contact offsets are written directly, per collision shape, through the asset's OvPhysxView: REST_OFFSET / CONTACT_OFFSET for articulations, RIGID_BODY_REST_OFFSET / RIGID_BODY_CONTACT_OFFSET for rigid objects. Both bindings are CPU-resident [N, S] buffers, so the full tensor is read-modify-written on the host with the selected envs as write indices.
• Dispatch now matches randomize_rigid_body_material in the same file: ovphysxmanager first (it contains the substring physx), then Newton, then PhysX, and a ValueError naming the manager for anything else instead of silently selecting PhysX.
• Docstring lists OVPhysX as a supported backend.

Note on write order

PhysX enforces restOffset < contactOffset and on OVPhysX a violating setRestOffset is only logged, not raised. The term writes rest before contact, same as the PhysX variant, so raising both above a small authored contact offset can silently skip the rest write. Kept for PhysX parity; flagging in case we want to reorder in both backends.

Verification

New source/isaaclab_ov/test/test_randomize_rigid_body_collider_offsets_mdp.py drives the public term (stubbed cfg / env / asset_cfg) against a real OVPhysX RigidObject and Articulation, so the dispatch itself is covered. It reproduced the reported AttributeError before the fix. Assertions: selected envs land in the sampled range, unselected envs are untouched, and omitting one distribution leaves that offset alone. Passes on cuda:0 and cpu (separate processes, ovphysx device lock). PhysX and Newton paths are unchanged.

Dependencies: #7563.

Fixes # (no tracking issue; reported by the OVQA agent harness)

Type of change

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

Release backport

• Backport this pull request to the active release branch after it merges into develop

Screenshots

Not applicable: no visual change.

Checklist

Docker and GPU tests run on demand. Push the commits you want tested, then
comment run-ci on the pull request.

• 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 added a changelog fragment under source/<pkg>/changelog.d/ for every touched package (do not edit CHANGELOG.rst or bump extension.toml — CI handles that)
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[greptile-apps]

Greptile Summary

Adds OVPhysX support to rigid-body collider-offset randomization and replaces implicit backend fallback with explicit dispatch.

• Adds an OVPhysX implementation using backend-specific per-shape tensor attributes.
• Routes OVPhysX, Newton, and PhysX managers explicitly and rejects unsupported managers.
• Adds real-backend tests for rigid objects and articulations on CPU and CUDA.
• Adds changelog fragments for the affected packages.

Confidence Score: 5/5

The PR appears safe to merge with no actionable correctness or security issues identified.

The backend dispatch now selects a dedicated OVPhysX implementation, and the added tests cover rigid-object and articulation writes for selected and all environments across CPU and CUDA configurations.

Important Files Changed

Filename	Overview
source/isaaclab/isaaclab/envs/mdp/events.py	Adds OVPhysX collider-offset handling and explicit backend selection while retaining the existing PhysX and Newton implementations.
source/isaaclab_ov/test/test_randomize_rigid_body_collider_offsets_mdp.py	Exercises public-term dispatch and selected/all-environment offset writes for real OVPhysX rigid objects and articulations on CPU and CUDA.
source/isaaclab/changelog.d/antoiner-fix-ovphysx-collider-offsets.minor.rst	Documents OVPhysX collider-offset support and explicit unsupported-backend errors.
source/isaaclab_ov/changelog.d/antoiner-fix-ovphysx-collider-offsets.skip	Marks the OVPhysX package change as test-only for changelog purposes.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[randomize_rigid_body_collider_offsets] --> B{Physics manager}
    B -->|ovphysxmanager| C[OVPhysX implementation]
    B -->|contains newton| D[Newton implementation]
    B -->|contains physx| E[PhysX implementation]
    B -->|other| F[Raise ValueError]
    C --> G[Read and write per-shape offset attributes]
    D --> H[Map offsets to Newton geometry properties]
    E --> I[Use PhysX root-view offset accessors]

Loading

_{Reviews (1): Last reviewed commit: "Support OVPhysX in randomize_rigid_body_..." | Re-trigger Greptile}

[isaaclab-review-bot]

Isaac Lab Review Bot

The OVPhysX backend and explicit dispatch extend randomize_rigid_body_collider_offsets as intended, but the new implementation can silently fail to apply a valid rest/contact target pair because it always writes the rest offset first.

• Design and architecture: The backend-specific class follows the existing dispatch and implementation pattern. The remaining concern is confined to OVPhysX write sequencing rather than the overall backend separation.
• API: The public term signature and existing PhysX/Newton paths remain unchanged, and OVPhysX support is documented. However, callers cannot rely on both requested offsets being applied when the target rest offset exceeds the currently authored contact offset, even if the final sampled pair is valid.
• Implementation: The CPU tensor routing, articulation-versus-rigid-object tensor selection, indexed full-buffer writes, and public-path tests support the implementation. The rest offset is nevertheless written before the contact offset; OVPhysX only logs rejection when the intermediate state violates rest < contact, so the rest write can be silently dropped. The writes should be sequenced according to current and target values so valid final pairs are applied reliably.

Minor fixes needed. Posted 1 actionable finding inline.

Automated review; human maintainers own approval decisions.

[isaaclab-review-bot]

🔵 Suggestion · Implementation — Rest offset written before contact offset

This new path writes the sampled rest offset before the contact offset. As the added test comment records, PhysX requires rest < contact and OVPhysX only logs a violating setRestOffset, so raising both above a small authored contact offset silently drops the rest write while contact succeeds, leaving a stale rest offset with no error. Sequencing the two writes based on current and target values (or writing contact first when it increases) would make every valid target pair land.

[kellyguo11]

run-ci

[isaaclab-bot]

Backported to release/3.0.0 as a31ae1a.