Fix OVPhysX device routing for CPU-resident tensor types

[AntoineRichard]

Description

_CPU_ONLY_TYPES in isaaclab_ov is the single constant deciding whether an OvPhysxView tensor buffer is allocated on, and validated against, host or device memory. It omitted eight tensor types that are CPU-resident even on a GPU simulation:

tensor type	alias	shape
ARTICULATION_CONTACT_OFFSET	CONTACT_OFFSET	[N, S]
ARTICULATION_REST_OFFSET	REST_OFFSET	[N, S]
RIGID_BODY_CONTACT_OFFSET	RIGID_BODY_CONTACT_OFFSET	[N, S]
RIGID_BODY_REST_OFFSET	RIGID_BODY_REST_OFFSET	[N, S]
ARTICULATION_BODY_DISABLE_GRAVITY	BODY_DISABLE_GRAVITY	[N, L]
RIGID_BODY_DISABLE_GRAVITY	RIGID_BODY_DISABLE_GRAVITY	[N]
ARTICULATION_DOF_DRIVE_TYPE	DOF_DRIVE_TYPE	[N, D]
ARTICULATION_DOF_DRIVE_MODEL	DOF_DRIVE_MODEL	[N, D, 3]

Because _native_device consults this set, it returned the simulation device for host-resident data. Two consequences:

1. A hidden per-call host-to-device staging copy on every access, on a path the view believed was device-native. Silent, and not attributable to Isaac Lab in a profile. Values are still correct — ovphysx stages the mismatch transparently, so this is not a data-integrity bug.
2. _check_device rejected a correctly placed buffer: allocating one of these on the host, where the data actually lives, raised OvPhysxView.DeviceMismatch, whose message advises moving the buffer yourself — precisely the advice that fails here.

No shipped Isaac Lab workflow reads these eight through the tensor path today, so this is latent rather than a live outage. It is not an obscure corner: Isaac Lab already authors these same properties through USD schema writes at scene construction, and the tensor path exists to change such values per-env at runtime.

Each of the eight is also exposed as a public alias carrying its shape, dtype and units, taken from the wheel's own type table. This module is the alias layer over the ovphysx enum, so referencing them raw would have left the candidate tuple a mix of aliases and raw enum members with no stated reason why these eight differed. Naming follows the existing convention: the ARTICULATION_ prefix is dropped for articulation types, as with DOF_STIFFNESS and BODY_MASS, and retained for rigid-body ones, as with RIGID_BODY_MASS.

A second, orthogonal gap surfaced during the audit: ovphysx documents ARTICULATION_DOF_DRIVE_TYPE as read-only, but it was missing from _READ_ONLY_NAMES, so set_attribute accepted writes and forwarded them. An audit of the full classification against the wheel's type table found this to be the only such gap — of the 21 types ovphysx documents read-only, 20 were already classified, with no misclassified members.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change — writes to articulation_dof_drive_type now raise OvPhysxView.ReadOnlyAttribute instead of being silently forwarded. Migration: remove the write; drive type is authored through the USD drive schema, not the tensor path.

Release backport

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

Screenshots

Not applicable.

Checklist

• 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

How it was tested

ovphysx exposes no residency query (ovphysx_tensor_spec_t carries dtype/ndim/shape and no device field), so residency was measured, not assumed: CUDA memcpys were counted via CUPTI around each binding read into a host buffer and into a device buffer, under the DirectGPU configuration isaaclab_ov itself sets ("/physics/suppressReadback": True). A host read showing DtoH means GPU-resident; no memcpy on the host read plus HtoD on the device read means CPU-resident.

Measured across every GPU sample scene in the ovphysx 0.5.11 wheel:

scene	before	after
two_articulations_gpu	agree=40 MISMATCH=8	agree=48 MISMATCH=0
links_chain_sample_gpu	agree=40 MISMATCH=8	agree=48 MISMATCH=0
mixed_base_articulations_gpu	agree=36 MISMATCH=8	agree=44 MISMATCH=0
boxes_falling_on_groundplane_gpu	agree=9 MISMATCH=3	agree=12 MISMATCH=0
volume_deformable_multi	MISMATCH=0	MISMATCH=0
surface_deformable_material	MISMATCH=0	MISMATCH=0

Every previously classified member measured CPU-resident as declared in all scenes, so the set was incomplete rather than wrong — there were no false positives to correct.

The guarding test was a tautology: it compared the view's derived set against the canonical set it is derived from, so it could not fail for any contents. It now asserts against an independent inventory of measured residency, mirroring the existing _EXPECTED_READ_ONLY_NAMES pattern. Both new assertions were confirmed to fail before the fix and pass after it; the read-only case fails with DID NOT RAISE ReadOnlyAttribute without the fix.

• pytest source/isaaclab_ov/test/sim/test_ovphysx_view.py — 81 passed
• test_articulation_helpers, test_rigid_object_helpers, test_deformable_object_helpers, test_ovphysx_scene_data_backend, test_randomize_rigid_body_material_mdp — all pass
• uv run isaaclab -f — passes
• tools/changelog/cli.py check develop — passes

test_views_xform_prim_ovphysx.py reports 23 failed / 28 passed both with and without this change; every failure is the process-global device lock covered by the device_split marker, which requires per-device re-invocation. Baselined on a clean tree to confirm it is unrelated.

Environment: ovphysx 0.5.11, Linux x86_64, Python 3.12, NVIDIA RTX 5000 Ada Generation Laptop GPU.

[greptile-apps]

Greptile Summary

This PR corrects OVPhysX tensor routing for eight CPU-resident types and classifies articulation DOF drive type as read-only.

• Adds the measured CPU-resident tensor types to the canonical CPU-only classification.
• Rejects unsupported writes to articulation_dof_drive_type.
• Replaces a tautological classification test with independent expected read-only and residency inventories.
• Documents the routing fix and intentional compatibility change.

Confidence Score: 5/5

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

The new classifications consistently feed the existing device-routing and read-only checks, are covered by independent inventory assertions, and the intentional behavior change is documented with migration guidance.

Important Files Changed

Filename	Overview
source/isaaclab_ov/isaaclab_ov/tensor_types.py	Adds eight measured host-resident OVPhysX tensor types to the CPU-only routing set; no actionable defect was found.
source/isaaclab_ov/isaaclab_ov/sim/views/ovphysx_view.py	Marks articulation DOF drive type read-only so unsupported writes fail before reaching OVPhysX.
source/isaaclab_ov/test/sim/test_ovphysx_view.py	Adds independent inventories that verify the complete CPU-only and read-only classifications.
source/isaaclab_ov/changelog.d/ovphysx-cpu-only-tensor-types.minor.rst	Clearly documents the corrected routing and migration required for the newly rejected write.

_{Reviews (1): Last reviewed commit: "Reject writes to the read-only DOF drive..." | Re-trigger Greptile}

[isaaclab-review-bot]

Isaac Lab Review Bot

The change updates the canonical CPU-residency classification for eight OVPhysX tensor types and marks articulation_dof_drive_type read-only, with focused inventory tests and migration guidance. The proposed changelog-version finding is not supported by a trusted rule establishing that this repository requires a major fragment for this contract change.

• Design and architecture: The reviewed design keeps device routing centralized in _CPU_ONLY_TYPES and derives the view’s name classification from that set, avoiding duplicate production tables. Using private _TT members avoids unnecessarily expanding the public tensor-type alias surface.
• API: No public symbols are added or removed. Writes to articulation_dof_drive_type now deterministically raise ReadOnlyAttribute, while correctly host-resident buffers for the eight added types pass device validation. The behavior change and migration path are explicitly documented.
• Implementation: The added classifications propagate through the existing read-only validation and native-device routing paths. The tests replace the prior self-derived CPU-only assertion with an independent measured inventory and extend the read-only inventory. The hard-coded inventory remains coupled to the supported OVPhysX wheel’s available enum members, which is a non-blocking compatibility tradeoff.

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.

[kellyguo11]

run-ci

[isaaclab-bot]

Backported to release/3.0.0 as f1f9576.