Physical artificial intelligence is moving from research labs into deployed robots and autonomous vehicles that must operate reliably in unpredictable environments, and Nvidia is positioning OpenUSD and Omniverse as the core infrastructure for building and testing these systems. The OpenUSD Core Specification 1.0 now defines common data types, file formats and composition behavior so developers can create predictable, interoperable pipelines for robotics and autonomous systems. On top of this standard, Omniverse libraries, Nvidia RTX rendering and physics simulation enable digital twins and SimReady assets, while Nvidia Cosmos world foundation models generate diverse weather, lighting and terrain conditions from the same scenes so teams can safely probe rare edge cases.
Nvidia is expanding the ecosystem around these standards to make simulation workflows more accessible and powerful. The Learn OpenUSD curriculum has been open sourced on GitHub, letting educators and teams localize and adapt training materials for OpenUSD-centric workflows. Techniques such as Gaussian splatting and new world models are speeding up high fidelity environment creation, highlighted by Nvidia Research’s Play4D 4D Gaussian splatting streaming pipeline and World Labs’ Marble generative world model, which can transform text prompts and sample images into photorealistic, physics ready environments in hours instead of weeks. Companies such as Lightwheel are building OpenUSD powered SimReady asset libraries with accurate geometry, materials and validated physical properties so robots trained in Isaac Sim and Isaac Lab experience realistic contacts, dynamics and sensor feedback.
The company is also tying these simulation advances to end to end safety frameworks for autonomous vehicles through Nvidia Halos. New research such as the Sim2Val framework combines real world and simulated test results to cut physical mileage requirements while still exposing robotaxis to rare, safety critical scenarios. An open source Omniverse NuRec Fixer model cleans artifacts in neural reconstructions to produce higher quality SimReady assets, and the Nvidia Halos Artificial Intelligence Systems Inspection Lab, accredited by ANAB, offers impartial inspection and certification of Halos elements across robotaxi fleets, full AV stacks, sensors and manufacturer platforms. Early participants include Bosch, Nuro and Wayve, while Onsemi is the first company to pass inspection. Ecosystem tools like the CARLA simulator, Voxel51’s FiftyOne and Mcity’s 32-acre AV test facility digital twin are integrating Omniverse, Cosmos and sensor modeling blueprints so developers can align real sensor recordings with high fidelity simulated data and share assets for safe, repeatable testing before vehicles reach public roads.
