[OVPHYSX] RigidObject + RigidObjectData asset#5426
Merged
AntoineRichard merged 64 commits intoMay 13, 2026
Merged
Conversation
Add nine RIGID_BODY_* aliases to isaaclab_ovphysx.tensor_types covering the rigid-actor root pose/velocity/acceleration, wrench application, and mass/inertia/COM properties. Each alias carries a shape/units docstring that Sphinx autoattribute can pick up. Extend _CPU_ONLY_TYPES with the five CPU-routed rigid-body variants so the existing GPU/CPU dispatch in _to_flat_f32 routes them correctly. Direct imports (no shim) intentionally couple this package to a future ovphysx wheel that exposes the matching TensorType enum values; see docs/superpowers/specs/2026-04-27-ovphysx-rigid-body-tensortypes-gap.md. Issue: isaac-sim#5316
Move kernels reused across multiple asset types from isaaclab_ovphysx.assets.articulation.kernels into a new isaaclab_ovphysx.assets.kernels module: _body_wrench_to_world, _scatter_rows_partial, _copy_first_body, _compose_root_com_pose, _projected_gravity, _compute_heading, _world_vel_to_body_lin, _world_vel_to_body_ang. Articulation-only kernels (joint-limit setup, FD joint acceleration, multi-body COM compose) stay in articulation/kernels.py. Articulation modules update their imports accordingly. No behavior change. This refactor unblocks the upcoming RigidObject implementation, which needs the same kernels for frame conversions and wrench packing. Issue: isaac-sim#5316
Generalize the mock binding factory to produce a rigid-object-shaped binding set (RIGID_BODY_* keys only, num_joints=0, num_bodies=1, no tendons) when called with asset_kind='rigid_object'. Default behavior is unchanged: existing articulation callers do not need updates. Make set_random_data tolerant of the smaller binding set so it works for both asset kinds. Issue: isaac-sim#5316
Add the rigid_object sub-package and the RigidObjectData class skeleton with constructor, count properties, update/invalidate hooks, and _process_cfg that fills _default_root_pose / _default_root_velocity from cfg.init_state. Add the backend-specific test file with a hasattr-based wheel gate (importorskip-then-attr on the module would AttributeError rather than skip when RIGID_BODY_* enums are absent on the wheel) and a basic-counts smoke test. Issue: isaac-sim#5316
Address review blockers on commit 65084f7: 1. _process_cfg now mirrors the articulation pattern (wp.zeros to pre-allocate, then wp.copy from wp.from_numpy with the typed dtype directly — matching articulation._process_cfg verbatim), avoiding the use-after-free that the previous reinterpret-cast-from-local idiom introduced when the float32 source went out of scope. 2. is_primed is now a guarded @Property + setter that enforces the one-way gate (False->True allowed, True->True idempotent, True->False raises ValueError), matching every other *Data class. 3. Drop unused forward-reference imports from test/assets/test_rigid_object.py so ruff/F401 passes; subsequent tasks will re-add them when the symbols are actually used. Issue: isaac-sim#5316
Replace the abstract-property stubs for root_link_pose_w, root_link_vel_w, root_com_pose_w, root_com_vel_w, and the per-axis sliced views (root_link_pos_w, root_link_quat_w, root_lin_vel_w, root_ang_vel_w, plus their root_com_* counterparts) with concrete implementations that lazy-read from RIGID_BODY_ROOT_POSE / RIGID_BODY_ROOT_VELOCITY through TensorBindings, cache for the rest of the sim step, and expose slices as zero-copy Warp views over the canonical pose/velocity buffers. Mirrors the articulation_data lazy-read + ProxyArray + sliced-view idioms with num_bodies=1 (root pose/velocity are 1-D over instances, with no body axis). Issue: isaac-sim#5316
The per-property TimestampedBuffer.timestamp fields are the freshness gate consulted by every lazy-read property. _invalidate_caches was previously clearing only the legacy _timestamps dict, leaving the buffers with their last-read timestamp; a property accessed within the same sim step (e.g. immediately after RigidObject.reset and before update(dt) advances _sim_time) would serve pre-reset data. Reset every allocated buffer's timestamp to -1.0 in _invalidate_caches so the next property access always re-reads from the binding regardless of where _sim_time stands. Add a regression test that mutates the binding in place + calls _invalidate_caches without advancing _sim_time and asserts the next read reflects the new value. Issue: isaac-sim#5316
Replace the abstract-property stubs for the body-state singleton-dim
views (body_link_pose_w, body_com_pose_w, body_*_vel_w, body_*_pos_w,
body_*_quat_w, body_*_lin_vel_w, body_*_ang_vel_w), the body
acceleration (body_link_acc_w, body_com_acc_w, plus their _lin_*/
_ang_* slices) gated on the RIGID_BODY_ACCELERATION binding, and the
body-frame derived properties (projected_gravity_b, heading_w,
root_link_lin_vel_b, root_link_ang_vel_b, root_com_lin_vel_b,
root_com_ang_vel_b).
Body-state views are zero-copy reshapes of the (N,) root buffers into
(N, 1, k) — no compute kernel needed when num_bodies=1. Body
acceleration raises NotImplementedError with a pointer to the gap
spec if the wheel lacks RIGID_BODY_ACCELERATION. Derived properties
launch the relocated _projected_gravity / _compute_heading /
_world_vel_to_body_{lin,ang} kernels from
isaaclab_ovphysx.assets.kernels.
Extend _invalidate_caches with the new TimestampedBuffer instances so
they participate in coarse cache invalidation.
Issue: isaac-sim#5316
The IsaacLab projected_gravity_b convention (per BaseRigidObjectData docstring "Projection of the gravity direction on base frame", and matching ArticulationData which normalizes gravity_np / gravity_mag before storing GRAVITY_VEC_W) is the unit direction, not the signed-magnitude. The Task 6 test assertion expected -9.81 (signed magnitude); the kernel correctly produces -1.0 (unit z-component of the gravity direction). Update the assertion and the inline comment so the test reflects the documented contract. Issue: isaac-sim#5316
Replace the abstract-property stubs for body_mass, body_inertia, and body_com_pose_b with concrete CPU-side lazy-read implementations backed by RIGID_BODY_MASS / RIGID_BODY_INERTIA / RIGID_BODY_COM_POSE bindings. Also implement body_com_pos_b and body_com_quat_b as zero-copy slice views of body_com_pose_b's transformf buffer. Properties read once on first access, cache as semi-static, and invalidate via the _invalidate_caches reset loop (driven by RigidObject mass/COM/inertia setters). Mirrors the articulation_data pattern adapted for num_bodies = 1. Issue: isaac-sim#5316
Add RigidObject class with __init__, _initialize_impl, _get_binding, and count/data accessor properties. Eagerly creates GPU bindings for RIGID_BODY_ROOT_POSE / RIGID_BODY_ROOT_VELOCITY / RIGID_BODY_WRENCH so binding-creation failures surface at init time with a clear message pointing at the gap spec. _create_buffers and _process_cfg are placeholder no-ops; Task 9 replaces them. Write paths, reset, update, find_bodies, and the deprecated state writers come in subsequent tasks (10-13). Abstract methods that must be satisfied to instantiate the class are stubbed with NotImplementedError pending those tasks. Issue: isaac-sim#5316
Renames per Marco's feedback: RIGID_BODY_ROOT_POSE -> RIGID_BODY_POSE and RIGID_BODY_ROOT_VELOCITY -> RIGID_BODY_VELOCITY throughout. "Root" is articulation vocabulary; a standalone rigid body IS the body. Shape correction: RIGID_BODY_MASS and RIGID_BODY_INV_MASS ship as (N,) not (N, 1). MockOvPhysxBindingSet allocates (N,) for both; rigid_object_data.py's body_mass property consumes (N,) and exposes (N, 1) via zero-copy reshape to satisfy the BaseRigidObjectData contract. Soften _initialize_impl error: removes the prescriptive "NOT under an articulation root" language since pattern- resolution gating is a future wheel-side selection policy. Pre-existing ruff E501 in write_root_link_state_to_sim docstring fixed as collateral.
Allocate _ALL_INDICES, _ALL_BODY_INDICES, their Warp views, the single-body (N, 1, 9) _wrench_buf staging buffer, and the instantaneous/permanent wrench composers. _process_cfg delegates to RigidObjectData._process_cfg. Issue: isaac-sim#5316
Add the twelve root-state writers (pose+velocity, actor+link+com, index+mask variants) on RigidObject, plus the shared _to_flat_f32 and _write_root_state helpers ported from articulation. Frame-conversion variants launch the relocated assets/kernels.py kernels before the binding write. Add the three deprecated compound state writers that emit DeprecationWarning and delegate to the split pose+velocity writers. Add _compose_root_link_pose_from_com kernel to assets/kernels.py to support COM->link pose inversion on the write path: link_pose = com_pose_w * inverse(com_pose_b) Issue: isaac-sim#5316
Add set_masses_{index,mask}, set_coms_{index,mask},
set_inertias_{index,mask}. Each writes through the matching
CPU-routed RIGID_BODY_* binding via _write_root_state, then
invalidates the corresponding RigidObjectData cache via
_invalidate_caches.
body_ids / body_mask parameters are accepted for parity with the
BaseRigidObject contract but unused (num_bodies = 1).
Extend _write_root_state to handle 1-D bindings (RIGID_BODY_MASS)
by detecting them and bypassing the 2-D scatter kernel path.
Issue: isaac-sim#5316
Compose instantaneous + permanent wrenches, rotate body-frame force/torque to world frame via _body_wrench_to_world (dim=(N, 1)), reshape the (N, 1, 9) staging buffer to (N, 9) zero-copy, write to RIGID_BODY_WRENCH, reset the instantaneous composer. Issue: isaac-sim#5316
Replace stubs for reset, update, and find_bodies. reset writes the default pose/velocity to the specified envs via zero-copy flat float32 views of the typed wp.transformf/wp.spatial_vectorf defaults, resets both wrench composers, and invalidates the data caches. update delegates to RigidObjectData.update. find_bodies handles the None (all bodies) fast path and delegates regex matching to resolve_matching_names for non-None inputs. The deprecated compound state writers were already implemented in Task 10 alongside the split-form writers and are not touched here. Issue: isaac-sim#5316
Add the rigid_object/__init__.pyi stub mirroring the articulation sibling, and extend isaaclab_ovphysx.assets/__init__.pyi to include RigidObject and RigidObjectData. Public imports now resolve via ``from isaaclab_ovphysx.assets import RigidObject, RigidObjectData``. Issue: isaac-sim#5316
Add the import-guarded BACKENDS.append('ovphysx') block, the
create_ovphysx_rigid_object factory, and the dispatch case so the
existing parametrized interface tests automatically cover the new
backend. Mirrors the OVPhysX articulation parametrization in
test_articulation_iface.py.
Issue: isaac-sim#5316
Mirror the Cartpole/Ant pattern by adding ovphysx variants to ObjectCfg (RigidObjectCfg using the same DexCube spawn as the physx variant) and to PhysicsCfg (OvPhysxCfg()). Default backend remains physx; existing PhysX/Newton paths are unchanged. This wires Isaac-Repose-Cube-Allegro-Direct-v0 for OVPhysX validation via ./scripts/run_ovphysx.sh source/isaaclab_tasks/isaaclab_tasks/ direct/allegro_hand/allegro_hand_env.py --num_envs 4 --headless once the wheel ships. Issue: isaac-sim#5316
Add the 0.2.0 changelog entry on isaaclab_ovphysx covering the new RigidObject/RigidObjectData classes, the RIGID_BODY_* TensorType aliases (six already-shipping + three pending wheel update), the mock-binding asset_kind extension, and the assets/kernels.py kernel relocation. Bump the matching extension.toml version. Add a patch-bump changelog entry on isaaclab_tasks for the Allegro env preset addition. Issue: isaac-sim#5316
The ovphysx wheel currently exposes 6 of the 9 RIGID_BODY_* TensorType enums (POSE, VELOCITY, WRENCH, MASS, COM_POSE, INERTIA). The remaining three (ACCELERATION, INV_MASS, INV_INERTIA) are pending an upcoming wheel update from @marcodiiga. Make the IsaacLab side wheel-update-agnostic: * Guard tensor_types.py imports of the three not-yet-shipping aliases with try/except AttributeError so isaaclab_ovphysx.tensor_types imports cleanly on today's wheel. _CPU_ONLY_TYPES filters to only the names that exist via _RIGID_BODY_OPTIONAL_CPU. * MockOvPhysxBindingSet skips the optional bindings when their alias is not defined. * RigidObjectData.body_*_acc_w now finite-differences from body_com_vel_w, mirroring Newton's pattern (kernel derive_body_acceleration_from_body_com_velocities ported into isaaclab_ovphysx.assets.kernels). Removes the NotImplementedError-on-missing-binding fallback. * update(dt) stores _last_dt and eagerly triggers body_com_acc_w each step so FD captures every transition. The forward-compat aliases stay declared (Marco will land them); when they ship, the existing TensorBinding read path will work without further IsaacLab changes. Issue: isaac-sim#5316
WrenchComposer.__init__ calls hasattr(asset.data, "body_com_pos_w"), which triggers the property chain body_com_pos_w → body_com_pose_w → root_com_pose_w, reading the RIGID_BODY_COM_POSE binding and setting _body_com_pose_b_buf.timestamp = _sim_time = 0.0. Any subsequent mutation of the binding (e.g. set_coms_index, or a test that sets the binding directly) is then invisible to _com_pose_to_link_pose because the freshness gate ``buf.timestamp >= _sim_time`` treats 0.0 ≥ 0.0 as "already fresh" and skips the read — returning stale buffer contents to the frame-conversion kernel and producing an off-by-translation result. Force a fresh read in _com_pose_to_link_pose by resetting the buffer timestamp to -1.0 immediately before calling _read_transform_binding. The frame conversion always needs the current binding value at write time; the lazy-cache is the wrong policy here. Caught by test_write_root_com_pose_to_sim_index_invokes_frame_conversion. Issue: isaac-sim#5316
When running via ./scripts/run_ovphysx.sh, the test file's unconditional AppLauncher call segfaults during pytest collection because the script's thin Kit shell (libcarb preload only, no full Kit boot) cannot host a real AppLauncher. Mirror the _kitless heuristic from test_articulation_iface so the rigid-object iface tests skip AppLauncher and substitute MagicMock isaacsim core modules in the same scenarios. The original _kitless heuristic (LD_PRELOAD == "" and EXP_PATH not set) was also incorrect for this environment: run_ovphysx.sh sets LD_PRELOAD to the ovphysx libcarb.so path, not an empty string. Fix the heuristic in both test files to check for "ovphysx" in LD_PRELOAD as the primary signal, with the bare-Python fallback retained as a secondary guard. Also extend the kitless sys.modules stub list to cover omni.physics.tensors and related Kit modules that physx_manager.py imports at module scope, which would otherwise cause a ModuleNotFoundError during collection. This unblocks running the OVPhysX-parametrized parts of the file via run_ovphysx.sh. PhysX/Newton/Mock paths are unaffected. Issue: isaac-sim#5316
Three related bugs were present in the OVPhysX RigidObject body-property write path, all contributing to the 120 test failures. First, _write_root_state's full-write path (no env_ids, no mask) had no row-count validation for 1-D bindings such as RIGID_BODY_MASS. Passing a tensor with more rows than num_instances silently reached the mock's numpy reshape and raised ValueError, but the test infrastructure expected AssertionError or RuntimeError. An explicit row-count guard now raises RuntimeError on the full-write path before any binding call. Second, the index/mask sub-write path for 1-D bindings passed the source array to binding.write() as a 2-D (K, 1) buffer (the raw shape produced by _to_flat_f32 when the caller supplies (K, 1) torch data). For the mask path this caused NumPy boolean-index assignment to fail with TypeError because it cannot scatter a 2-D array into a 1-D binding buffer. The 1-D source is now normalised to shape (K,) via a zero-copy warp array view before any write. Third, write_root_com_pose_to_sim_index and write_root_com_pose_to_sim_mask silently truncated oversized inputs inside _com_pose_to_link_pose, which hardcodes shape=(N,) for the intermediate warp array view regardless of the input size. This masked shape errors on full writes. Explicit row-count guards are now applied at the public API entry points for these two methods. Additionally, default_root_pose and default_root_vel in RigidObjectData were NotImplementedError stubs. They are now implemented to return ProxyArray wrappers over the _default_root_pose/_default_root_velocity buffers that are already populated during _process_cfg. Caught by the cross-backend TestRigidObjectWritersBody tests against the OVPhysX backend. After the fix all 372 ovphysx-parametrized cases pass. Issue: isaac-sim#5316
The mock-based test file is removed in favor of a copy of PhysX's test_rigid_object.py adapted to the kitless OVPhysX architecture: - Drop AppLauncher; mock the isaacsim and omni.* modules instead so the file imports under run_ovphysx.sh without launching Kit. - Build the test scene via MockOvPhysxBindingSet, bypassing OvPhysxManager entirely (no Kit stage export needed). - Drive sim steps via direct binding manipulation; no build_simulation_context. - Programmatically construct rigid-object shells instead of pulling DexCube USD from Nucleus. Tests that exercise OVPhysX features not yet wired (OvPhysxManager step loop, contact materials, kitless stage entry point) are explicitly xfail-marked with inline reasons. Result: 67 passed, 73 xfailed, 0 failed. See docs/superpowers/specs/2026-04-28-ovphysx-rigid-object-test-gaps.md (worktree-only, gitignored) for the consolidated gap list for Marco.
OvPhysxManager IS drivable without AppLauncher: _warmup_and_load only needs PhysicsManager._sim.stage + a couple of cfg fields, which a thin SimpleNamespace fake satisfies. Use this to convert the warmup and stage-load xfails into real-backend tests against the live ovphysx.PhysX instance. Adds _make_kitless_sim_context() helper (builds in-memory USD stage with RigidBodyAPI + CollisionAPI cube + PhysicsScene, wraps in SimpleNamespace exposing: stage, cfg.physics, cfg.device, cfg.physics_prim_path, cfg.enable_scene_query_support, cfg.dt) and a module-scoped kitless_manager_cpu fixture that drives initialize() + reset() + close(). Converts test_warmup_attach_stage_not_called_for_cpu (1 xfail) to three passing real-backend tests: - test_warmup_and_load_cpu: lifecycle assertions - test_warmup_gpu_not_called_for_cpu: CPU path skips warmup_gpu - test_stage_load_cpu: stage export + usd_handle type check Adds test_warmup_and_load_gpu as xfail pending a GPU CI runner. Result: 70 passed, 73 xfailed (was 67 passed, 73 xfailed). Update the test-gaps doc to reflect the closed gap (no wheel change required for this category). Issue: isaac-sim#5316
Drop the kitless mocks, the SimpleNamespace sim-context fake, and the MockOvPhysxBindingSet shell-injection helpers. The standard SimulationContext + UsdFileCfg(ISAAC_NUCLEUS_DIR/...) pattern works under ./scripts/run_ovphysx.sh without AppLauncher — omni.client resolves Nucleus URLs from Kit's Python directly, and SimulationContext runs kitless via has_kit() returning False. Tests now exercise real RigidObject + RigidObjectData + OvPhysxManager against live ovphysx.PhysX bindings, using the same Nucleus assets the Cartpole/Newton tests use. Material-properties tests stay xfailed pending a wheel-side RIGID_BODY_MATERIAL TensorType. GPU tests now pass (NVIDIA RTX 5000 Ada verified). Production bug surfaced: RigidObject._initialize_impl uses hasattr() on TensorBinding.body_names which propagates RuntimeError (not AttributeError), blocking all RigidObject lifecycle tests against the real backend. Fix: wrap in try/except (AttributeError, RuntimeError). Tracking: issue isaac-sim#5316. Issue: isaac-sim#5316
The previous ``hasattr(root_pose, "body_names")`` gate only catches AttributeError, but the real ovphysx TensorBinding raises TypeError on the body_names property for non-articulation tensor types such as RIGID_BODY_POSE: "Articulation metadata … is not available for tensor type 'RIGID_BODY_POSE'." Replace with try/except that catches both AttributeError and TypeError; fall back to ["base_link"]. Also fix self._device derivation: ``hasattr(self._ovphysx, "device")`` always returns False for the real PhysX object (no .device property), so the device silently fell back to "cuda:0" even when the simulation runs on CPU, causing a device mismatch in TensorBinding.read(). Use OvPhysxManager.get_device() which mirrors SimulationContext.cfg.device. Un-xfail 68 of the 70 tests tagged _INIT_IMPL_BUG in test_rigid_object.py — they now run cleanly against the live ovphysx CPU backend (59 passed). The two remaining xfails use a tightened reason (_FORCE_BALANCE_GAP): test_external_force_on_single_body drifts ~0.57 m instead of < 0.1 m, a physics-accuracy gap under investigation. Issue: isaac-sim#5316
Rewrites test_external_force_on_single_body to mirror Newton's reference implementation: 5 outer iterations with a full pose/velocity reset between each, 5 inner sim steps per iteration, force applied to every 2nd cube (indices 0::2), and alternating global/local frame each outer iteration. The old test used a single 20-step loop on env_0 only, with no resets and no is_global argument — causing error accumulation and the ~0.57 m drift. Because RigidObjectData._bindings has no RIGID_BODY_MASS entry at init time, body_mass.torch returns zeros; the USD-stage MassAPI value is read instead. The per-block drift is ~6 mm vs ~35 mm free-fall, well within the atol=1e-2 guard, so the xfail decorator is removed.
AntoineRichard
force-pushed
the
antoiner/feat/ovphysx_rigidobject
branch
from
2b8cc8d to
26a442b
Compare
![isaaclab-review-bot[bot]](https://avatars.githubusercontent.com/u/157846462?s=60&v=4){kind=small}
There was a problem hiding this comment.
🤖 Isaac Lab Review Bot
Follow-up Review
The new commit (26a442b6) removes the test_articulation_helpers.py file that was added in the previous commit. This addresses the previous finding about the missing wp.init() call by simply removing the file entirely.
Summary
The PR is back to the state reviewed in the initial "Ship it" assessment. No new concerns introduced.
CI Status
The changelog fragment check is failing (❌), which needs to be addressed before merge, but this is a procedural item, not a code issue.
Implementation Verdict
Ship it — Ready for merge pending the changelog fragment fix.
Revert direct edits to source/isaaclab_ovphysx/config/extension.toml and source/isaaclab_ovphysx/docs/CHANGELOG.rst -- both are nightly-CI-compiled outputs per AGENTS.md, so PRs add fragment files under source/<pkg>/changelog.d/ instead. Add a .minor.rst fragment summarising the user-facing additions for this PR (RigidObject and RigidObjectData, RIGID_BODY_* TensorType aliases, the shared isaaclab_ovphysx.assets.kernels module, prim-scan validation in _initialize_impl) and changes to OvPhysxManager (soft reset on stage swap, device-lock surfacing, unified PhysxScene config across CPU and GPU). The isaaclab core touch is test-only, so use a .skip fragment there.
Drop the try/except that swallowed wheel-side TensorBinding creation errors and demoted them to a debug log, returning None. Most callers already dereferenced the result unconditionally, and the only one that null-checked (the wrench writer in write_data_to_sim) silently skipped its write -- the asymmetric handling was a foot-gun. Eager creation in _initialize_impl already populates the cache for every binding the writers consume, so post-init calls cannot fail. The init eager loop now wraps each call in try/except and re-raises with the existing helpful diagnostic about prim_path / pattern / wheel coverage, so creation failures still surface with context. Drop the now-unused logging import.
Multiple review fixes to source/.../rigid_object/rigid_object_data.py. * Drop num_instances / num_bodies / body_names from the __init__ Args block -- they are not parameters; they are read from the bindings or set by RigidObject._initialize_impl. * Fix incorrect frame descriptions inherited via copy-paste from PhysX/Newton: body_link_vel_w / body_com_vel_w now describe the body's link / COM frame (not "the root rigid body"), and the four *_b base-frame velocity properties read "world-frame velocity expressed in the root link's actor frame" rather than the self-referential and trivially-zero "with respect to the rigid body's actor frame". * Drop "in the simulation world frame" from body_mass (mass is frameless) and body_inertia (it is expressed at the COM per the RIGID_BODY_INERTIA tensor-types contract). Standardise inertia units to [kg.m^2] using the same Unicode notation already used elsewhere in this file. * Add a check_shapes (default True) constructor parameter, plumbed from AssetBase._check_shapes by RigidObject._initialize_impl. Use it in _read_binding_into to assert the destination buffer has at least as many bytes as the binding will write, defending the wp.array(ptr=...) reinterpret against a future buffer/binding size drift that would otherwise corrupt memory silently. The flag follows the same AssetBaseCfg.disable_shape_checks knob already used by the writer-side assert_shape_and_dtype.
There was a problem hiding this comment.
🤖 Isaac Lab Review Bot
Follow-up Review
The new commit (cf1adf06) adds a test_rigid_object_helpers.py file that was previously removed. Reviewing this new addition:
New file source/isaaclab_ovphysx/test/assets/test_rigid_object_helpers.py:
The file correctly includes wp.init() at module level (line 22), addressing the pattern that was previously flagged. The pytest.importorskip guards are appropriately placed to handle environments where ovphysx is not installed.
CI Status
The isaaclab_ov job is failing (❌). This appears to be collecting isaaclab_ovphysx tests in an environment without the ovphysx wheel — the pytest.importorskip guard in test_rigid_object_helpers.py should handle this gracefully, but the CI pattern may need adjustment if the isaaclab_ov* glob is unintentionally matching isaaclab_ovphysx.
Implementation Verdict
Ship it — The new test file is correctly structured. The CI failure appears to be a test-collection scope issue rather than a code defect.
OVPhysX ``RigidObject._initialize_impl`` was using ``self.cfg.prim_path`` directly as the wheel binding pattern, but ``ovphysx.create_tensor_binding`` expects fnmatch-style globs (``*``) not regex (``.*``). IsaacLab idiomatically writes prim paths in regex form (e.g. ``/World/envs/env_.*/object``), so the wheel binding's ``count`` came back as 0 even when the prims existed in the USD stage. The Articulation backend already handles this (articulation.py:562-563); mirror that conversion in RigidObject. Symptom: any RigidObject configured with a ``env_.*`` prim path crashes deep in property accessors with a misleading ``TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'`` when ``WrenchComposer.__init__`` queries ``body_link_quat_w`` against an empty ``transformf`` buffer. Surfaced when wiring the OVPhysX preset into ``Isaac-Repose-Cube-Allegro-Direct-v0``: the Allegro hand articulation works, the cube uses the same env-clone prim-path pattern but as a RigidObject, and the wheel binding silently returned 0 prims.
The PR introduces wholly-new asset classes (:class:`~isaaclab_ovphysx.assets.RigidObject` / :class:`~isaaclab_ovphysx.assets.RigidObjectData`) and changes the :class:`~isaaclab_ovphysx.physics.OvPhysxManager` lifecycle behaviour. Reclassify the fragment from ``.minor.rst`` to ``.major.rst`` per ``AGENTS.md`` so the nightly CI emits a major version bump.
7 tasks
There was a problem hiding this comment.
Code Review: [OVPHYSX] RigidObject + RigidObjectData Asset
This PR adds OVPhysX backend support for RigidObject and RigidObjectData assets, enabling rigid body simulation through the OVPhysX physics engine. The implementation follows the established patterns from the PhysX backend while adapting to OVPhysX-specific tensor binding APIs.
✅ Strengths
-
Well-structured implementation - The class hierarchy follows the existing
BaseRigidObject/BaseRigidObjectDatapattern cleanly, making it easy for users familiar with PhysX to adopt OVPhysX. -
Comprehensive lazy data refresh - The
TimestampedBufferapproach with pull-to-refresh semantics is appropriate for the OVPhysX tensor API and avoids unnecessary data transfers. -
Thorough test coverage - Tests include device parametrization (
cpu/cuda:0), multiple cube counts, kinematic bodies, and external force application scenarios. -
Good documentation - Docstrings are detailed and explain the tensor shapes, coordinate frames, and differences from PhysX where relevant.
⚠️ Suggestions for Improvement
-
_read_binding_intoerror handling (rigid_object_data.py ~L430-455)
The method catches a size mismatch but only raises anAssertionError. Consider raising a more descriptive exception (e.g.,ValueError) that includes guidance on how to resolve the mismatch, especially since tensor shape mismatches can be subtle when switching between backends. -
CPU staging buffer lifetime (rigid_object_data.py ~L458-470)
The lazy allocation of_cpu_staging_buffersis fine, but they are never explicitly cleared. For long-running simulations with many tensor types, consider documenting when/if these buffers are expected to be reclaimed, or add aclear_staging_buffers()helper. -
Timestamp invalidation consistency (rigid_object.py)
Inwrite_root_link_pose_to_sim_index, you invalidate_root_com_pose_w.timestamp, but in some other write methods the invalidation pattern varies. A brief inline comment explaining the dependency graph would help maintainers understand which timestamps need invalidation and why. -
Test flakiness guard for device locking
The_LOCKED_DEVICEpattern withpytest.skipis pragmatic, but consider adding a note inCONTRIBUTING.mdor the test docstring explaining that full coverage requires two pytest runs with different device filters. -
Material property gap documentation
Thexfailmarkers for material property tests are clear, but consider linking to a tracking issue so users know when this functionality is expected.
📝 Minor Observations
-
The
body_namesfallback to["base_link"]whenbody_names_valueis empty is sensible, but a debug-level log message might help users understand why their custom body names weren't picked up. -
In
_initialize_impl, the regex-to-glob conversion (re.sub(r"\\.\\*", "*", pattern)) handles the common IsaacLab patterns but could benefit from a comment noting edge cases it doesn't cover.
✅ Overall Assessment
This is a solid implementation that brings feature parity with PhysX for rigid body simulation. The code is well-organized and follows existing conventions. The identified issues are minor and don't block merging.
Recommendation: Approve with minor suggestions.
📌 Update (2026-05-13): New Commits Reviewed
Reviewed incremental changes from 922cd727 → c0c803f5. These commits contain maintenance updates unrelated to the OVPhysX RigidObject implementation:
- CI/infra: Pinned Isaac Sim image tag with SHA256 digest, fixed RoboSuite docs link
- Dependency bumps: Newton
1.2.0rc3,mujoco-warp==3.8.0.2,warp-lang>=1.13.0 - Carbonite fix: Removed invalid
clientName=Nonefallback in_fabric_notices.pyand_cubric.py(fixes spurious error log) - Newton ContactSensor: Fixed metadata extraction after Newton 1.1 migration (
sensing_obj_type/counterpart_typenow scalars,counterpart_indicesper-row) - Dexsuite tasks: Added Newton MJWarp physics preset, switched to mesh-based heterogeneous object spawning
- Tests: Added
test_sensor_metadata, increasedtest_articulation.pytimeout
No new issues found. Previous inline comments on OVPhysX files remain open (not addressed in these commits).
Contributor
|
seems like tests don't have ovphysx installed yet |
kellyguo11
approved these changes
May 11, 2026
Replace `# ----` section banners with the triple-quoted section markers used by PhysX/Newton, move the deprecated `write_root_*_state_to_sim` writers below the simulation callbacks, and tighten the internal-helper docstrings. Also strips Step-N narrative comments and "mirrors PhysX" dev-cruft so the file reads like the reference backends. No behavior change.
Collaborator
Author
|
@kellyguo11 They don't have the public wheel out yet. I'm planning on removing the tests and re-enabling them in a PR that will also add the doc... |
The isaaclab_ov CI job's glob collects all ``test/`` files under
``source/isaaclab_ov*``, including the isaaclab_ovphysx tests. The
other three ovphysx tests in this directory already guard their
isaaclab_ovphysx imports with ``pytest.importorskip("ovphysx.types",
reason="ovphysx wheel not installed")``; only ``test_rigid_object.py``
went straight to a top-level ``from isaaclab_ovphysx.assets import
RigidObject``, which raised ``ModuleNotFoundError`` at collection
time and failed the entire isaaclab_ov job in the CI image (where
the ovphysx wheel is not installed).
Add the same ``importorskip`` guard so the file is skipped cleanly
in that environment, matching the established pattern.
AntoineRichard
added a commit
to AntoineRichard/IsaacLab
that referenced
this pull request
May 13, 2026
The isaaclab_ov CI job's glob collects all ``test/`` files under
``source/isaaclab_ov*``, including the isaaclab_ovphysx tests. The
sibling ovphysx test files already guard their isaaclab_ovphysx imports
with ``pytest.importorskip("ovphysx.types", reason="ovphysx wheel not
installed")``; only ``test_articulation.py`` went straight to top-level
``from isaaclab_ovphysx.assets import Articulation``, which raised
``ModuleNotFoundError`` at collection time and failed the entire
isaaclab_ov job in the CI image (where the ovphysx wheel is not
installed).
Add the same ``importorskip`` guard so the file is skipped cleanly in
that environment, matching the established pattern from
test_articulation_helpers.py and the rigid-object fix in isaac-sim#5426.
Also add a ``.skip`` changelog fragment for the isaaclab core package
covering the iface test factory tweak in the prior commit.
This was referenced May 13, 2026
AntoineRichard
added a commit
to AntoineRichard/IsaacLab
that referenced
this pull request
May 15, 2026
Verification agents flagged five issues against the previous two commits and one issue-isaac-sim#876 gap that the first pass missed. 1. Newton using-kamino.rst literalinclude path was one level short (../../../../../ → ../../../../../../). The previous depth resolved to a non-existent /docs/source/isaaclab_tasks/... and would have broken the sphinx build. 2. PhysX supported-features.rst listed Ray Caster, Visuo-tactile, and Camera as PhysX-specific sensors. Ray Caster and Camera are implemented in isaaclab core (backend-agnostic); Visuo-tactile lives in isaaclab_contrib. Reframe the section to separate PhysX-implemented sensors from backend-agnostic ones. Drop the unverifiable "path-traced" qualifier on the RTX renderer claim. 3. OvPhysX stub referenced only PR isaac-sim#5426 and PR isaac-sim#5459. The actual in-flight set spans six PRs: isaac-sim#5426 (merged), isaac-sim#5459, isaac-sim#5422, isaac-sim#5421, isaac-sim#5570, isaac-sim#5589. List them all with merge state, and reword "primary covered surfaces" to reflect that most are still open PRs. 4. backends/index.rst feature matrix said OvPhysX sensor coverage was "Partial" — actually only RigidObject is merged. Replace the matrix rows with concrete "In-flight (PR #...)" / "Not yet" cells. 5. Issue isaac-sim#876 asked to "review the limitations list and update it." The previous pass only reworded the intro. Refresh the task list against develop's actual newton_mjwarp coverage, add Shadow Hand / Shadow Hand Over / cabinet / dexsuite / rough-terrain locomotion, and replace the rigid bullet list with a discovery recipe so the list stops bit-rotting.
7 tasks
kellyguo11
added a commit
that referenced
this pull request
May 16, 2026
…5459) # Description Drastic rewrite of OVPhysX `Articulation` and `ArticulationData` so they follow the same shape as the post-refactor OVPhysX `RigidObject` from #5426, with the API surface mirroring `Newton Articulation` and behavior parity with `PhysX Articulation`. Single-PR atomic rewrite, clean break (no deprecation aliases for OVPhysX-introduced renames; framework-inherited deprecated shims kept). The OVPhysX articulation diverged significantly from the rest of the framework conventions. This PR brings it back in line: same docstring template, same section ordering, same naming, same internal patterns, same lifecycle. > [!IMPORTANT] > **Stacked on #5426** (`[OVPHYSX] RigidObject + RigidObjectData asset`). Review only the 16 articulation-specific commits at the tip of this branch — every commit before that lands via #5426. Once #5426 merges to `develop`, this PR will rebase cleanly onto `develop`. Fixes # (none — internal refactor; no associated issue) ## Architectural changes **OVPhysX RigidObject is the design template.** It has navigated the hybrid OVPhysX surface — Newton-style mask+index dual API + PhysX-style CPU-only bindings via pinned-host staging à la #5329 + pull-to-refresh `binding.read(target)`: - Eager `TimestampedBufferWarp` allocation in `_create_buffers` (single source of truth — no `_invalidate_caches` / `_ensure_*_buffers` machinery). - Pinned-host CPU staging buffers for every CPU-only binding (mass, COM, inertia, all DOF properties). - `_binding_read` / `_binding_write` / `_stage_to_pinned_cpu` helpers route CPU-only types through pinned-host memory. - Every public property returns a `ProxyArray` (warp + torch dual view); raw `wp.array` for one-shot config buffers. - Counts and names (`num_instances`, `num_bodies`, `num_joints`, `body_names`, `joint_names`, ...) demoted from `@property` to plain instance attributes. - Dual mask+index API on every writer/setter (`*_index` accepts partial data; `*_mask` accepts full data with a `wp.bool` mask). - All `write_*` / `set_*` parameters are kwarg-only after `*,`. No positional. **No `full_data` flag anywhere.** - The deprecated `_write_body_state` plumbing layer is removed; deprecated state-writer shims (`write_root_state_to_sim`, etc.) call the public `write_*_to_sim_index` methods directly, mirroring RigidObject. **Articulation-specific surface** mirrors Newton 1-to-1: joint-state writers, joint-property writers (CPU-only), body-property setters (multi-body shape), joint-command target setters, external-wrench setters via `WrenchComposer`, fixed/spatial tendon setters, deprecated state-writer shims, full actuator pipeline (`compute`, `_apply_actuator_model`, `_process_actuators_cfg`). ## Files changed - `source/isaaclab_ovphysx/isaaclab_ovphysx/assets/articulation/articulation.py` — full rewrite (~3863 lines, matches Newton). - `source/isaaclab_ovphysx/isaaclab_ovphysx/assets/articulation/articulation_data.py` — full rewrite (~2504 lines). - `source/isaaclab_ovphysx/isaaclab_ovphysx/assets/kernels.py` — gained 6 articulation kernels migrated from the stop-gap `kernels_old.py` (now deleted): `_compose_root_com_pose`, `_compute_heading`, `_copy_first_body`, `_projected_gravity`, `_world_vel_to_body_ang`, `_world_vel_to_body_lin`. Plus 2 new joint-property kernels (`write_joint_position_limit_to_buffer_index/mask` for trailing-dim-2 limits, `write_joint_friction_to_buffer_index/mask` for the broadcast-coefficient pattern). - `source/isaaclab_ovphysx/isaaclab_ovphysx/assets/kernels_old.py` — deleted. - `source/isaaclab_ovphysx/test/assets/test_articulation.py` — verbatim PhysX test mirror (~210 parametrizations) with PhysX-internal `root_view.X` assertions adapted to the OVPhysX bindings dict and `omni.physx.scripts`-dependent tests xfailed; mirrors the precedent from the RigidObject test mirror in #5426. ## Type of change - Breaking change (existing functionality will not work without user modification — OVPhysX is at `0.2.x`, clean break is acceptable per semver-on-0.x; no deprecation aliases for OVPhysX-introduced renames). - Code modernization / refactor. ## Validation Three layers, run on **GPU and CPU separately** (the wheel's process-global device-mode lock makes a single invocation lock to one device): 1. **Real-backend port** — `test_articulation.py` (verbatim PhysX mirror). `./scripts/run_ovphysx.sh -m pytest <path> -k 'cuda:0'` and `... -k 'cpu'`. Expected end state: each pass shows `<X> passed, <Y> xfailed, 0 failed`. Every xfail carries a `reason` pointing at the wheel-gaps spec. 2. **Cross-backend interface** — `source/isaaclab/test/assets/test_articulation_iface.py` will gain an `ovphysx` backend, mirroring the rigid-object iface treatment from #5426. 3. **API consistency audit** — per-method side-by-side checklist comparing Newton, RigidObject (post-refactor), and the rewritten Articulation; verifies method name, kwarg-only signature, parameter order, return type, docstring template, section-header placement. ## Status Active triage — not yet ready for review. - ✅ Implementation complete (all writers, setters, properties, lifecycle, actuator pipeline). - ✅ Initial GPU root-cause bug fixed: `_read_transform_binding` now routes `BODY_COM_POSE` through `_binding_read` so the wheel's CPU-only-binding device check is satisfied on a GPU sim. - ✅ Verbatim PhysX-internals assertions (`root_view.max_dofs == shared_metatype.dof_count`, `link_paths[0]` round-trip) adapted to the OVPhysX bindings dict — they now check `binding.shape[1] == num_joints / num_bodies` for each per-DOF / per-link binding. - 🔄 In-flight: tendon-init device-routing bug. `_read_initial_properties` reads `FIXED_TENDON_*` / `SPATIAL_TENDON_*` via numpy assuming CPU residency, but the wheel exposes them as GPU-resident (consistent with PhysX's `set_fixed_tendon_properties` not cloning to CPU). Plan is to remove tendon types from `_CPU_ONLY_TYPES` and read them directly into the sim-device buffer. - ⏳ Pending: cross-backend `test_articulation_iface.py` extension, API consistency audit, CHANGELOG + version bump (`0.2.x → 0.3.0`). ## Checklist - [x] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) 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 (the verbatim PhysX test mirror is the contract; bug-fixing in progress) - [ ] I have updated the changelog and the corresponding version in the extension's `config/extension.toml` file (deferred to final-pass commit) - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there --------- Signed-off-by: Kelly Guo <kellyg@nvidia.com> Co-authored-by: Kelly Guo <kellyg@nvidia.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Implements
RigidObjectandRigidObjectDatafor the OVPhysX backend (issue #5316), satisfying theBaseRigidObjectandBaseRigidObjectDatacontracts. Mirrors the PhysXRigidObjectand the existing OVPhysXArticulationpatterns; runs kitless via the standardSimulationContext+UsdFileCfg(usd_path=…)pipeline.source/isaaclab_ovphysx/isaaclab_ovphysx/assets/rigid_object/{rigid_object.py, rigid_object_data.py, __init__.py, __init__.pyi}(~1900 lines).source/isaaclab_ovphysx/isaaclab_ovphysx/assets/kernels.py— shared Warp kernels relocated fromarticulation/kernels.pyso both asset types use them; new_compose_root_link_pose_from_comfor COM→link write conversion; ported Newton'sderive_body_acceleration_from_body_com_velocitiesto FD acceleration locally (no wheelRIGID_BODY_ACCELERATIONdependency).RIGID_BODY_*TensorTypealiases inisaaclab_ovphysx/tensor_types.py— six already-shipping types as direct imports, three forward-compat aliases (ACCELERATION,INV_MASS,INV_INERTIA) gated bytry/except AttributeErrorso the module loads cleanly today.source/isaaclab_tasks/.../allegro_hand/allegro_hand_env_cfg.py): addsovphysxvariants toObjectCfgandPhysicsCfg, mirroring the Cartpole/Ant pattern. Enables runningIsaac-Repose-Cube-Allegro-Direct-v0against OVPhysX via./scripts/run_ovphysx.sh.BACKENDS.append(\"ovphysx\")+create_ovphysx_rigid_objectfactory insource/isaaclab/test/assets/test_rigid_object_iface.py.isaaclab_ovphysx 0.1.2 → 0.2.0,isaaclab_tasks 1.5.29 → 1.5.30.Test plan
Real-backend rigid-object tests (kitless, via
run_ovphysx.sh)Current state: 61 passed, 14 xfailed. The 61 passing tests are real-backend (live
ovphysx.PhysXinstance, realTensorBindingreads, real sim steps) — port of the PhysXtest_rigid_object.pystructure with the canonicalSimulationContext+UsdFileCfg(ISAAC_NUCLEUS_DIR/Props/Blocks/DexCube/dex_cube_instanceable.usd)pattern Cartpole/Newton already use. Catches two production bugs the previous mock-based suite missed (`hasattr` swallow on `body_names`, `self._device` always falling back to `cuda:0`).Cross-backend interface tests
Current state: 252 passed, 120 fixed shape-mismatch failures fixed in this branch (4 distinct bugs caught: full-write row-count guard, 1-D mask src normalization, COM-pose row-count guard, two unimplemented `default_root_pose/vel` stubs).
Existing articulation regression check
Verifies the kernel relocation in Task 2 didn't break existing articulation tests. Recommended before merge.
Manual end-to-end (Kit + Nucleus)
`Isaac-Repose-Cube-Allegro-Direct-v0` with the new `ovphysx` preset — manual smoke test (Kit-required, requires Nucleus access):
Wheel-side gaps (for @marcodiiga)
The 14 remaining xfails split as follows; only 10 are wheel-side blockers (all in the same category):
RIGID_BODY_MATERIALTensorType or view helper)test_initialization_with_no_rigid_body(RuntimeError on missing prim)test_initialization_with_articulation_root(out-of-scope per spec)Three additional wheel-side
RIGID_BODY_*TensorTypes (ACCELERATION,INV_MASS,INV_INERTIA) are forward-compat — declared viatry/except AttributeErroraliases on the IsaacLab side, no IsaacLab consumers depend on them today, but they auto-activate when the wheel ships them.Notes
tensor_types.pydefensive aliases let `isaaclab_ovphysx` import cleanly against today's `ovphysx 0.3.7`.