Fix OVPhysX device routing for CPU-resident tensor types

_CPU_ONLY_TYPES omitted eight tensor types that are CPU-resident even on a
GPU simulation: the articulation and rigid-body collision-shape contact and
rest offsets, both gravity-disable flags, and the DOF drive type and drive
model.

Because _native_device consults this set, it returned the simulation device
for host-resident data. Every access staged a hidden host-to-device copy on a
path the view believed was device-native, and _check_device rejected a buffer
placed on the host, where the data actually lives, with DeviceMismatch.

Residency was measured rather than assumed: ovphysx exposes no residency
query, so each type was classified by counting CUDA memcpys around a binding
read into a host buffer and into a device buffer, under the DirectGPU
configuration isaaclab_ov itself sets. All eight measured CPU-resident across
every GPU sample scene in the ovphysx wheel, and all previously classified
members measured as declared.

The guarding test 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, so an incomplete set is caught.

Author: AntoineRichard

source/isaaclab_ov/changelog.d/ovphysx-cpu-only-tensor-types.minor.rst

@@ -0,0 +1,11 @@
+Fixed
+^^^^^
+
+* Fixed :class:`~isaaclab_ov.sim.views.OvPhysxView` routing eight CPU-resident
+  tensor types to the simulation device. The per-collision-shape contact and rest
+  offsets, the articulation and rigid-body gravity-disable flags, and the DOF drive
+  type and drive model are CPU-resident even on a GPU simulation, but were absent
+  from the internal CPU-only classification. Reads and writes of these types
+  incurred a hidden per-call host-to-device staging copy, and a correctly placed
+  host buffer was rejected with ``OvPhysxView.DeviceMismatch``. Residency was
+  measured on a GPU simulation by counting CUDA memcpys around a binding read.

source/isaaclab_ov/isaaclab_ov/tensor_types.py

@@ -442,6 +442,17 @@
     RIGID_BODY_MASS,
     RIGID_BODY_COM_POSE,
     RIGID_BODY_INERTIA,
+    # Collision-shape offsets, gravity-disable flags and DOF drive metadata. Measured
+    # CPU-resident on a GPU sim; referenced through ``_TT`` since this module exposes
+    # no public alias for them.
+    _TT.ARTICULATION_DOF_DRIVE_TYPE,
+    _TT.ARTICULATION_DOF_DRIVE_MODEL,
+    _TT.ARTICULATION_BODY_DISABLE_GRAVITY,
+    _TT.ARTICULATION_CONTACT_OFFSET,
+    _TT.ARTICULATION_REST_OFFSET,
+    _TT.RIGID_BODY_DISABLE_GRAVITY,
+    _TT.RIGID_BODY_CONTACT_OFFSET,
+    _TT.RIGID_BODY_REST_OFFSET,
     DEFORMABLE_MATERIAL_DYNAMIC_FRICTION,
     DEFORMABLE_MATERIAL_YOUNGS_MODULUS,
     DEFORMABLE_MATERIAL_POISSONS_RATIO,

source/isaaclab_ov/test/sim/test_ovphysx_view.py

@@ -70,6 +70,48 @@
     }
 )
 
+# CPU-resident on a GPU sim. Each name was measured, not assumed: with DirectGPU
+# (``suppressReadback=True``) the residency of every tensor type was determined by counting
+# CUDA memcpys around a binding read into a host buffer and into a device buffer.
+_EXPECTED_CPU_ONLY_NAMES = frozenset(
+    {
+        "articulation_body_com_pose",
+        "articulation_body_disable_gravity",
+        "articulation_body_inertia",
+        "articulation_body_inv_inertia",
+        "articulation_body_inv_mass",
+        "articulation_body_mass",
+        "articulation_contact_offset",
+        "articulation_dof_armature",
+        "articulation_dof_damping",
+        "articulation_dof_drive_model",
+        "articulation_dof_drive_type",
+        "articulation_dof_friction_properties",
+        "articulation_dof_limit",
+        "articulation_dof_max_force",
+        "articulation_dof_max_velocity",
+        "articulation_dof_stiffness",
+        "articulation_rest_offset",
+        "articulation_shape_friction_and_restitution",
+        "deformable_material_bending_damping",
+        "deformable_material_bending_stiffness",
+        "deformable_material_dynamic_friction",
+        "deformable_material_elasticity_damping",
+        "deformable_material_poissons_ratio",
+        "deformable_material_thickness",
+        "deformable_material_youngs_modulus",
+        "rigid_body_com_pose",
+        "rigid_body_contact_offset",
+        "rigid_body_disable_gravity",
+        "rigid_body_inertia",
+        "rigid_body_inv_inertia",
+        "rigid_body_inv_mass",
+        "rigid_body_mass",
+        "rigid_body_rest_offset",
+        "rigid_body_shape_friction_and_restitution",
+    }
+)
+
 # Per-type shapes used by the fakes (only the types touched by the tests).
 _SHAPES = {
     TensorType.RIGID_BODY_POSE: lambda n: (n, 7),
@@ -204,15 +246,21 @@ def test_read_only_and_cpu_only_classification():
     assert not is_read_only("articulation_dof_stiffness")
     assert is_cpu_only("articulation_dof_stiffness")
     assert is_cpu_only("rigid_body_mass")
+    assert is_cpu_only("rigid_body_disable_gravity")
+    assert is_cpu_only("articulation_dof_drive_type")
     assert not is_cpu_only("rigid_body_pose")
 
 
 def test_cpu_only_names_match_canonical_set():
-    # The view derives its CPU-only set from tensor_types so the two cannot drift.
+    # Comparing the view's derived set against the canonical one cannot fail, since the
+    # former is built from the latter. The expected inventory is what gives this test
+    # teeth: it is measured residency, so it also rejects a set that is merely incomplete.
     from isaaclab_ov.sim.views import ovphysx_view as mod
     from isaaclab_ov.tensor_types import _CPU_ONLY_TYPES
 
-    assert frozenset(tt.name.lower() for tt in _CPU_ONLY_TYPES) == mod._CPU_ONLY_NAMES
+    assert frozenset(tt.name.lower() for tt in _CPU_ONLY_TYPES) == _EXPECTED_CPU_ONLY_NAMES
+    assert mod._CPU_ONLY_NAMES == _EXPECTED_CPU_ONLY_NAMES
+    assert set(attribute_vocabulary()) >= mod._CPU_ONLY_NAMES
 
 
 # -----------------------------------------------------------------------------