OvPhysX

OvPhysX runs PhysX without Omniverse Kit. It reads scene-level solver settings from the USD PhysicsScene prim. OvPhysxCfg holds GPU buffer settings that USD does not represent.

OvPhysX is selected through OvPhysxCfg:

from isaaclab.sim import SimulationCfg
from isaaclab_ov.physics import OvPhysxCfg
sim_cfg = SimulationCfg(physics=OvPhysxCfg())

Why use OvPhysX?

• No Kit runtime. Run headless simulations without Isaac Sim.
• USD configuration. Scene-authoring tools can set solver parameters directly on the PhysicsScene prim.

What works today

OvPhysX supports the following components. The notes describe requirements and limits to consider when configuring a task.

See /source/concepts/sensors/index for sensor configuration and /source/concepts/renderers for renderer choices.

Implementation history: rigid objects #5426, articulations #5459, rigid-object collections #5570, contacts #5422, scene data #5589, and frame views #5678 in https://github.com/isaac-sim/IsaacLab/pulls. Broader integration work was tracked in issue #5634. Check isaaclab_ov implementations and tests when updating this table.

Installation

The ovphysx extra requires OvPhysX 0.6.4.72118723. Install it from the repository root with:

uv sync --inexact --extra ovphysx

The --inexact flag preserves packages installed through other extras. Use --extra ov to install both public OvPhysX and OVRTX runtimes. The combined extra pairs OVRTX 0.5.1.385782 with OVStage 0.2.1; OVRTX 0.4.1 is not compatible with this runtime combination. The legacy Isaac Lab installer also supports ./isaaclab.sh -i 'ov[ovphysx]' and ./isaaclab.sh -i 'ov[all]'.

Testing the Installation

First check that the Python package and runtime wheel import correctly:

uv run --extra ovphysx --extra test python -c "import ovphysx.types; from isaaclab_ov.physics import OvPhysxCfg; print('OvPhysX runtime OK')"

Then run a small backend smoke test:

uv run --extra ovphysx --extra test python -m pytest source/isaaclab_ov/test/assets/test_rigid_object.py::test_initialization -k cpu

To try a task that declares an OvPhysX physics preset, use the same preset CLI syntax as the other backends:

uv run --extra ovphysx isaaclab zero_agent --task Isaac-Cartpole-Direct \
    --num_envs 128 --max_steps 64 --viz none physics=ovphysx

This command runs a 64-step headless zero-action rollout and then exits.

Supported locomotion environments

Use the environment browser or task help for the available physics= presets. To list launchable combinations:

uv run --extra ovphysx isaaclab list_envs --show_presets --keyword Velocity

Playback uses the same physics=ovphysx selector as training.

Heterogeneous cloning

OvPhysX 0.6.3 is required for heterogeneous runtime cloning. Tensor bindings use numeric environment order so indexed resets, actions and observations address the correct variant. Variants may differ in geometry but must preserve body, joint and tendon layout. Isaac Lab checks prototype rotation axes and tendon layouts; OVPhysX checks body and joint metadata when binding articulations.

The clone context compiles the plan’s prototypes and world assignments once. The physics manager attaches one exported stage, replays the native copies, and warms the runtime before assets and sensors bind. Hard reset reuses the same declarations; there is no separate consumable queue. Binding paths come from the plan’s world layout and environment template, not from native clone order.

Authored sources all receive native environment ID zero. Every world containing a source or USD-only physics therefore imports all its declared assets as originals. Other worlds clone their assets from a complete original world of the same composition, using one native environment ID per destination world. This keeps a robot, table and object able to contact one another without destination placeholders. Only declared asset subtrees are copied; unrelated authored assets are retained.

Keep filter_collisions enabled to isolate original worlds with USD collision groups. GPU copies additionally use native environment-ID filtering. CPU has no native environment-ID filtering in OvPhysX 0.6.3, so it imports all declared USD copies and uses collision groups. Large overlapping original layouts can still exhaust broadphase pair capacity; use spatially separated environments.

Deformable limitations

OvPhysX deformables currently require every body matched by one DeformableObject to have the same number of simulation nodes. Initialization raises an error for mixed node counts instead of exposing padded state that would produce incorrect reductions.

Deformable scenes also require full-stage materialization. Startup cost therefore grows with the number of authored environments. Use this path for small validation scenes; training-scale workloads with thousands of environments are not currently supported. Deformable bodies are not supported by the heterogeneous fast-path cloner.

Native access

Use native access when the shared asset and sensor APIs do not expose the data or operation you need. Reuse an Isaac Lab-owned handle when its selection matches; create raw access only for additional capabilities. Native writes can bypass Isaac Lab’s caches and validation, so follow the ownership and synchronization rules below.

Mental model

OvPhysX exposes generic TensorBinding objects. A binding combines a prim selection with one installed TensorType; it is not an asset-specific native storage object. Callers explicitly pull values into their own buffer with read() and push a buffer back with write(). Isaac Lab optionally wraps these bindings in OvPhysxView for a string-keyed, guarded convenience surface.

Lifecycle prerequisite

Create bindings only after the OvPhysX simulation has initialized and reset, when its native runtime is live. A stage reload, a reset path that rebuilds the stage or runtime, and manager teardown invalidate all existing bindings and views. Reacquire them afterwards.

Get the active OvPhysX handle from the manager after the lifecycle prerequisite:

from isaaclab_ov.physics import OvPhysxManager
physx = OvPhysxManager.get_physx_instance()
if physx is None:
    raise RuntimeError("OvPhysX has not been constructed; initialize and reset the simulation first.")

A non-None handle proves only that the manager constructed an OvPhysX instance; it is not a readiness predicate for the current simulation context. Manager teardown releases the instance and clears this handle. Initialize and reset the current context before access, and always reacquire the handle before creating new bindings or views.

Reuse Isaac Lab-owned access

Reuse the convenience view already owned by an Isaac Lab asset when its selection matches:

object_view = scene["object"].root_view
poses = object_view.get_attribute("rigid_body_pose")

Create raw access

Create a raw binding when the desired selection is not already owned by an Isaac Lab asset. The following representative pose binding is float32 and uses the simulation device; it is not a generic allocator for every tensor type:

import warp as wp
from isaaclab.physics import PhysicsManager
from ovphysx.types import TensorType
binding = physx.create_tensor_binding(
    pattern="/World/envs/env_*/Object",
    tensor_type=TensorType.RIGID_BODY_POSE,
)
try:
    poses = wp.empty(
        tuple(binding.shape),
        dtype=wp.float32,
        device=PhysicsManager.get_device(),
    )
    binding.read(poses)
    binding.write(poses)
finally:
    binding.destroy()

Before allocating a buffer, inspect binding.shape, binding.dtype, the binding’s count, and its path metadata. Match the binding shape and DLPack scalar metadata exactly, and respect the binding’s native device and access mode. RIGID_BODY_POSE has the float32 layout used above, but another tensor type can have a different scalar dtype, shape, native device, or write permission.

For a new selection, construct OvPhysxView with the native runtime and a Tensor API pattern:

from isaaclab.physics import PhysicsManager
from isaaclab_ov.sim.views import OvPhysxView
view = OvPhysxView(
    physx,
    pattern="/World/envs/env_*/Object",
    device=PhysicsManager.get_device(),
)
try:
    if view.try_binding_for("rigid_body_pose") is not None:
        poses = view.get_attribute("rigid_body_pose")
        view.set_attribute("rigid_body_pose", poses)
        view.read_into("rigid_body_pose", poses)
finally:
    view.close()

try_binding_for attempts to create a binding for this selection and returns None when a valid type is unavailable. binding_for returns a raw binding and bypasses the view’s device, dtype, shape, and read-only guards.

Read/write semantics

Raw binding.read(buffer) calls pull values into caller-owned buffers, while binding.write(buffer) calls push values to the simulation. Editing a local buffer alone does not change the simulation. The guarded get_attribute, read_into, and set_attribute paths provide the same pull/push boundary while validating the binding’s shape, scalar dtype, native device, and writable access.

Ownership, synchronization, and invalidation

Allocate buffers on the binding’s required device and match its shape and DLPack scalar metadata. State tensors normally use the simulation device, while some property tensors are CPU-only. OvPhysxView validates these requirements and can reinterpret supported flat scalar layouts as structured Warp values, but it does not move data between CPU and GPU.

Use try_binding_for when a tensor type might not apply to the selected prims. Reacquire raw bindings and convenience views after reset paths that rebuild the stage or runtime, or after teardown. Destroy raw bindings and close custom views before they are no longer needed. Manager teardown closes tracked OvPhysxView instances before releasing the runtime, but explicit cleanup avoids retaining their bindings for the rest of a long-lived simulation.

Authoritative references

The generated Isaac Lab reference for OvPhysxView documents the convenience view. The repository metadata pins OvPhysX to a version but does not provide a versioned official OvPhysX documentation URL, so inspect the installed TensorType and binding metadata at runtime.

Where to go next

• /source/isaaclab/environments — tasks and backend presets.
• Backend architecture — shared factories and resource ownership.
• /source/concepts/scene_data_providers — shared scene data for renderers and visualizers.