> This page is for version Release 3.0 (default).
> For other versions, use one of these documentation indexes:
> - Release 3.0 (default): https://docs.nvidia.com/isaaclab/3.0/llms.txt
> - Release 2.3: https://docs.nvidia.com/isaaclab/2.3/llms.txt
> - Release 2.2: https://docs.nvidia.com/isaaclab/2.2/llms.txt
> - Release 2.1: https://docs.nvidia.com/isaaclab/2.1/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/isaaclab/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/isaaclab/_mcp/server.

# OvPhysX Backend

> **Warning**
>
> OvPhysX is **highly experimental** and is not recommended for general use yet.
> The public surface is changing rapidly while the backend is under active
> development. Expect feature coverage and test commands to change between
> Isaac Lab 3.0 beta releases.

> **Warning**
>
> Do not combine OvPhysX with the Kit visualizer. Commands such as
> `physics=ovphysx --visualizer kit` are unsupported because ovphysx
> loads USD-dependent PhysX plugins from its own package, while Kit already
> owns a separate USD/plugin stack in the same process. Use
> `--visualizer newton_gl`, `--visualizer rerun`, `--visualizer viser`,
> or omit `--visualizer` for headless execution.

OvPhysX is a kit-less variant of the PhysX backend. It drives PhysX directly
(without the Omniverse Kit runtime) and reads scene-level solver parameters
from the USD `PhysicsScene` prim rather than from a Python config. The Python
config `~isaaclab_ov.physics.OvPhysxCfg` only exposes the handful of
GPU buffer sizes that are not represented on the USD schema.

OvPhysX is selected through `~isaaclab_ov.physics.OvPhysxCfg`:

```python
from isaaclab.sim import SimulationCfg
from isaaclab_ov.physics import OvPhysxCfg

sim_cfg = SimulationCfg(physics=OvPhysxCfg())
```

## Why use OvPhysX?

* **Kit-less execution.** OvPhysX avoids Omniverse Kit, which makes it a useful
  experimental path for headless deployments and for backends that don't need
  the Kit runtime stack.
* **USD-as-source-of-truth.** Solver parameters are taken from the
  `PhysicsScene` USD prim, so authoring tools that already manage USD scenes
  do not need a parallel Python config.

## What works today

The asset and sensor surface tracks PhysX, but only a subset is implemented and
validated at the time of writing. The following pieces are available on
`develop`:

* RigidObject — merged via
  [PR #5426](https://github.com/isaac-sim/IsaacLab/pull/5426).
* Articulation — merged via
  [PR #5459](https://github.com/isaac-sim/IsaacLab/pull/5459).
* RigidObjectCollection — merged via
  [PR #5570](https://github.com/isaac-sim/IsaacLab/pull/5570).
* Contact Sensor — merged via
  [PR #5422](https://github.com/isaac-sim/IsaacLab/pull/5422).
* SceneDataProvider — merged via
  [PR #5589](https://github.com/isaac-sim/IsaacLab/pull/5589).
* FrameView — merged via
  [PR #5678](https://github.com/isaac-sim/IsaacLab/pull/5678).
* `~isaaclab.assets.DeformableObject` — experimental volume- and
  surface-deformable support on CUDA simulation devices.
* Fast-path cloning of heterogeneous rigid-body and articulation geometry
  variants with matching native tensor layouts.

Additional OvPhysX work remains in flight. IMU, Frame Transformer, Joint Wrench,
PVA, Ray Caster, and rendering support are not documented as supported here
until their implementations land on `develop` and pass the backend smoke
tests.

## 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 `isaaclab.scene.InteractiveSceneCfg.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
`~isaaclab.assets.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.

## Installation

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

```bash
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 (Recommended)

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

#### isaaclab.sh / isaaclab.bat

```bash
./isaaclab.sh -p -c "import ovphysx.types; from isaaclab_ov.physics import OvPhysxCfg; print('OvPhysX runtime OK')"
```

Then run a small backend smoke test:

#### uv (Recommended)

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

#### isaaclab.sh / isaaclab.bat

```bash
./isaaclab.sh -p -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 (Recommended)

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

#### isaaclab.sh / isaaclab.bat

```bash
./isaaclab.sh -p scripts/environments/zero_agent.py --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

The following locomotion training environments declare an `ovphysx` physics
preset. Playback through `uv run isaaclab play` uses the same backend where
available.

* `Isaac-Ant-Direct`
* `Isaac-Humanoid-Direct`
* `IsaacContrib-Velocity-Rough-AnymalB`
* `IsaacContrib-Velocity-Rough-AnymalC`
* `Isaac-Velocity-Flat-AnymalD`
* `Isaac-Velocity-Rough-AnymalD`
* `IsaacContrib-Velocity-Rough-UnitreeA1`
* `IsaacContrib-Velocity-Rough-UnitreeGo1`
* `Isaac-Velocity-Rough-UnitreeGo2`
* `Isaac-Velocity-Rough-Cassie`
* `Isaac-Velocity-Rough-G1`
* `Isaac-Velocity-Rough-H1`

## Status and follow-up

OvPhysX is still experimental, so the feature list above is intentionally
conservative. Broader feature coverage and documentation parity are tracked in
[issue #5634](https://github.com/isaac-sim/IsaacLab/issues/5634).

For architectural context, see [Backend architecture](/source/concepts/backends_and_presets#backend-architecture).

## Native access

### 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 `~isaaclab_ov.sim.views.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:

```python
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:

```python
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:

```python
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 `~isaaclab_ov.sim.views.OvPhysxView`
with the native runtime and a Tensor API pattern:

```python
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
`~isaaclab_ov.sim.views.OvPhysxView.get_attribute`,
`~isaaclab_ov.sim.views.OvPhysxView.read_into`, and
`~isaaclab_ov.sim.views.OvPhysxView.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. `~isaaclab_ov.sim.views.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
`~isaaclab_ov.sim.views.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.