OpenUSD and NVIDIA Halos are driving major advances in the safety and scalability of physical AI systems, particularly in autonomous vehicles and robotaxis. As these technologies move from research labs into real-world deployment, developers face the challenge of ensuring reliability in unpredictable environments. The OpenUSD Core Specification 1.0 now provides a standardized foundation for data types, file formats, and composition behaviors, enabling consistent, interoperable workflows across tools and teams. Built on OpenUSD, NVIDIA Omniverse offers a powerful ecosystem for creating digital twins and simulation-ready (SimReady) assets. These assets include precise geometry, materials, and validated physical properties, allowing robots and AVs to experience realistic dynamics, sensor feedback, and environmental interactions in virtual environments. NVIDIA RTX rendering and physics simulation further enhance fidelity, making simulations highly accurate representations of the real world. A key innovation is the use of generative world models like NVIDIA Cosmos, which can create diverse and complex scenarios from a single base scene. By generating variations in weather, lighting, and terrain, these models help teams test edge cases safely and efficiently. Techniques like 4D Gaussian splatting, demonstrated through NVIDIA’s Play4D pipeline, enable fast, high-fidelity rendering of dynamic environments. Companies like World Labs are leveraging these capabilities to turn text prompts and images into photorealistic, physics-ready 3D worlds in hours rather than weeks. Lightwheel’s SimReady asset library, powered by OpenUSD, simplifies the creation of digital twins for robotics. These assets are designed for immediate use in NVIDIA Isaac Sim and Isaac Lab, accelerating robot training with realistic physics and sensor behavior. For autonomous vehicles, safety validation is being transformed through frameworks like Sim2Val, developed by NVIDIA researchers with Harvard and Stanford. This approach statistically combines real-world and simulated test data, reducing the need for massive physical test miles while still proving safety in rare, high-risk situations. The open-source NVIDIA Omniverse NuRec Fixer, trained on AV data, improves the quality of neural reconstructions by removing artifacts, further enhancing simulation accuracy. The NVIDIA Halos AI Systems Inspection Lab, accredited by ANAB, provides independent certification of AI systems across robotaxi fleets, AV stacks, sensors, and platforms. Early adopters include Bosch, Nuro, Wayve, and Onsemi, the first company to pass inspection. The lab ensures that systems meet rigorous safety and performance standards. Open tools are also playing a key role. The CARLA simulator now integrates NVIDIA NuRec and Cosmos Transfer for high-quality scene reconstruction and scenario generation. Voxel51’s FiftyOne engine, connected to Cosmos Dataset Search, helps teams curate, annotate, and evaluate multimodal datasets. At Mcity, the University of Michigan is using Omniverse to build a digital twin of its 32-acre test facility, incorporating physics-based sensor models to enable safe, repeatable testing of dangerous driving scenarios. The Learn OpenUSD curriculum is now open source on GitHub, offering educators and teams a flexible, community-driven way to adopt OpenUSD workflows. These advancements, combined with the power of OpenUSD and NVIDIA’s ecosystem, are paving the way for safer, faster, and more scalable deployment of physical AI systems across industries.