NVIDIA has added Omniverse libraries to its Agent Toolkit, giving software developers callable components for building AI agents that inspect 3D scenes, simulate sensors and physics, and prepare digital assets for robotics and industrial simulation. The new libraries and a reference integration for Blender are openly available on GitHub.
The update is aimed at a practical bottleneck in physical AI. A visually convincing 3D model is not automatically suitable for simulation: it also needs correct scale, structure, materials, labels, sensors and physical properties. NVIDIA says agents can now help check those requirements and prepare assets inside tools that designers and engineers already use.
Three core Omniverse capabilities
The ovrtx library provides RTX-based sensor simulation. Applications can generate synthetic camera, lidar, radar and other sensor outputs from a 3D scene, allowing developers to test how a robot or autonomous system may perceive a virtual environment.
The ovphysx library adds GPU-accelerated physics. It models properties such as collision, mass, friction and motion so an agent or developer can evaluate how objects behave before a system is exposed to them in the real world.
CAD-to-SimReady skills convert computer-aided design data into assets based on OpenUSD with properties needed for simulation and virtual testing. Together, the components are intended to let an agent inspect a scene, propose or apply changes and validate whether an asset is ready for downstream physical-AI work.
Integrations with established 3D tools
SideFX is exploring the libraries in Houdini workflows, including the use of OpenUSD, ovrtx and ovphysx for procedural 3D content. PTC is using OpenUSD and ovrtx with its Onshape CAD and product-data-management platform to connect design content with simulation.
NVIDIA also names ForgeCAD, Lightwheel, Moonlake AI and Palatial as early adopters. Their work ranges from automatically converting CAD inputs into validated assets to generating physically accurate simulation content from text prompts.
A new SimReady Blender blueprint shows how software makers can embed sensor simulation, physics and validation in an existing application. NVIDIA says the demonstration uses its NemoClaw software and can run locally, from compact RTX Spark systems through to GB300-powered DGX Station systems.
Availability across local and cloud systems
The Omniverse libraries, CAD-to-SimReady skills and Blender reference are available through GitHub. NVIDIA’s announcement emphasises support for both cloud and local AI systems, allowing developers to choose an environment based on model size, scene complexity, data restrictions and performance requirements.
RTX Spark systems are scheduled to become available in the northern autumn from ASUS, Dell Technologies, HP, Lenovo, Microsoft Surface and MSI, with Acer and Gigabyte models expected later. DGX Station systems are already available to order from several hardware partners. Those hardware details are relevant to the reference deployment but are not prerequisites for inspecting the open libraries.
What the update changes
Agent Toolkit already connects AI agents with tools, skills and data sources. The Omniverse addition extends that pattern into 3D and simulation work, where an agent may need to reason about both digital content and physical constraints.
The components may reduce repetitive preparation and validation work, but developers remain responsible for checking the resulting assets and simulations. A virtual environment can omit conditions that matter in deployment, and an agent’s proposed correction may be syntactically valid while still being physically or operationally wrong.
NVIDIA has not published comparative performance figures or detailed licensing terms in the announcement. Teams should review the relevant repositories, documentation and dependency licences before using the libraries in production software.
A practical adoption path
Software teams can begin with the Blender blueprint or a contained validation task rather than handing an agent control of an entire content pipeline. For example, an agent could flag missing collision properties or inconsistent units and leave the correction for a technical artist to approve. That produces measurable time savings while keeping responsibility with a domain expert.
As confidence grows, the same tools could be allowed to generate sensor configurations, convert larger CAD libraries or run batches of physics checks. Logs should record the source asset, agent instruction, library version and resulting changes so a simulation problem can be traced back to its preparation step.
Compatibility will also matter. OpenUSD improves interchange between applications, but real production pipelines include custom schemas, plug-ins and naming conventions. Early adopters should test those details before assuming that an asset validated in one environment will behave identically in every simulator or design tool.
The open release model should make those tests easier because teams can inspect the libraries and build repeatable checks around them. It also creates version-management work: an asset prepared with one library release may need revalidation after physics, sensor or schema updates. Pinning dependencies and recording validation results will be important for regulated or safety-critical workflows.