Skip to content

ovphysx

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

ovphysx

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
include/ovphysx
include/ovphysx
 
 
licenses
licenses
 
 
python
python
 
 
tests
tests
 
 
LICENSE.txt
LICENSE.txt
 
 
README.md
README.md
 
 
VERSION
VERSION
 
 

README.md

ovphysx

PyPI Python 3.10+ Linux | Windows

1. What is ovphysx?

ovphysx is a self-contained library for USD-based physics simulation, exposing a C API with Python bindings. It wraps NVIDIA PhysX SDK and loads USD scenes directly, runs the simulation, and exchanges state with your code via DLPack tensors � all with no Omniverse Kit installation required. It packages its own OV-namespaced OpenUSD runtime, so you get USD-native physics as a standalone dependency you can drop into any Python or C/C++ application.

NVIDIA PhysX SDK is the popular and stable real-time physics simulation engine underneath � the same GPU-accelerated rigid-body, articulation, and contact solver used across robotics, simulation, and interactive 3D for over a decade. PhysX is open source under BSD-3-Clause and ships in the same repository alongside the Omniverse PhysX extensions (for Kit-based apps). ovphysx is the path that brings that engine to developers as a lightweight, USD-first, kitless library.

2. What functionalities are available, and who are the target users?

What you can do with it:

  • Load and simulate USD scenes - add_usd("scene.usda"), step(dt, ...), release(). A full sim loop in three calls.
  • Tensor data exchange via DLPack - read/write simulation state (e.g. rigid-body poses) as zero-copy tensors that interoperate with NumPy, PyTorch, and other ML frameworks.
  • Tensor bindings - bind to scene patterns (e.g. RIGID_BODY_POSE) to stream state in and out efficiently.
  • Environment cloning for batched RL - replicate environments for high-throughput reinforcement-learning workloads.
  • CPU or GPU simulation - run with an NVIDIA GPU for acceleration, or fall back to CPU-only.
  • Standalone C/C++ SDK - pre-built packages on GitHub Releases with ready-to-build samples (find_package(ovphysx)): hello world, tensor bindings, cloning, contacts, and OmniPVD recording for debugging.

Who benefits:

  • Robotics & RL developers (Isaac Lab / Isaac Sim style workflows) who need fast, batched, headless physics with direct tensor access for policy training and inference.
  • Simulation & digital-twin engineers who want USD-native physics without pulling in a full Omniverse/Kit stack.
  • C/C++ application developers integrating real-time physics into non-Python engines and tools via the standalone SDK.
  • Researchers and ML practitioners who live in NumPy/PyTorch and want simulation state as plain tensors.

3. Documentation and reference links

4. System requirements

  • Python 3.10+
  • Linux (x86_64, aarch64) or Windows (x86_64)
  • NVIDIA GPU + driver recommended for acceleration (CPU-only simulation also supported)
  • Current release: ovphysx 0.5.9, compatible with PhysX SDK 5.10.0

5. Licensing

  • Source code: BSD-3-Clause License - permissive, free for commercial and non-commercial use.
  • Pre-built binaries (SDK packages and Python wheels): distributed under the NVIDIA Omniverse License.

Note: ovphysx is pre-release and not yet mature. When sharing a process with other OV USD-aware subsystems, register each subsystem's schema paths before the first USD stage or schema-registry access. Parts of the API may change before 1.0.

Note: Pre-release notice: ovphysx is pre-release software and not yet mature. ovphysx packages an OV namespaced OpenUSD runtime; when sharing a process with other OV USD-aware subsystems, register each subsystem's schema paths before the first USD stage or schema-registry access. Parts of the API are still being completed and may change before 1.0.

Quick Start

pip install ovphysx
from ovphysx import PhysX

physx = PhysX()
physx.add_usd("scene.usda")
physx.step(1.0 / 60.0, 0.0)
physx.release()

Environment Cloning

Clone environments for batched reinforcement-learning workloads:

from ovphysx import PhysX
from ovphysx.types import TensorType
import numpy as np

physx = PhysX(device="cpu")
usd_handle, _ = physx.add_usd("scene.usda")
physx.wait_all()

# Read rigid body poses via DLPack-compatible tensors
pose_binding = physx.create_tensor_binding(
    pattern="/World/envs/env0/table",
    tensor_type=TensorType.RIGID_BODY_POSE,
)
poses = np.zeros(pose_binding.shape, dtype=np.float32)
pose_binding.read(poses)

pose_binding.destroy()
physx.remove_usd(usd_handle)
physx.release()

C/C++ SDK

A standalone C SDK is available for integration into non-Python applications. Download pre-built packages from the GitHub Releases page.

After extracting the SDK, you can build and run a bundled sample directly:

# /path/to/ovphysx-sdk is the extracted SDK package (pre-built binaries from GitHub Releases)
cmake -B build -S /path/to/ovphysx-sdk/samples/c_samples/hello_world_c -DCMAKE_PREFIX_PATH=/path/to/ovphysx-sdk
cmake --build build
./build/hello_world_c

The SDK includes ready-to-build samples in samples/c_samples/ covering core workflows (hello world, tensor bindings, cloning, contacts, OmniPVD recording). Each sample has its own CMakeLists.txt that uses find_package(ovphysx).

Documentation

Links