NVIDIA Unveils Open Physical AI Models and Tools for Robotics, Autonomous Vehicles, and Industrial Automation

NVIDIA Research breakthroughs demonstrate that training across diverse gripper types, driving scenarios, and virtual worlds produces AI that generalizes to applications it has never encountered. A robot gripper's usefulness depends not on picking up one object, but on handling the next object, and the one after that, with tools it has never held before. This scaling principle applies equally to autonomous vehicles, where safety requires reasoning fast enough on hardware actually installed in production cars.

The research emphasizes exposure to maximum environmental diversity before deployment. Virtual agents trained across varied simulated worlds transfer skills more reliably to physical systems. This approach addresses the fundamental challenge in physical AI: the gap between simulation and reality that has historically limited robot deployment.

NVIDIA's Cosmos 3 platform powers these new physical AI agent skills, accelerating data generation, simulation, policy training, and evaluation for autonomous system development. The company unveiled these capabilities at CVPR, the premier computer vision conference.

NVIDIA released a major collection of open-source physical AI skills and tools spanning its Omniverse, Cosmos, Alpamayo, and Metropolis platforms. These tools transform complex physical AI training, evaluation, and deployment workflows into repeatable, optimized, and agent-executable instructions. The release includes Alpamayo-R1, described as the world's first open industry-scale reasoning vision language action model for mobility applications.

The Cosmos and GR00T open models provide robot learning and reasoning capabilities, while Isaac Lab-Arena offers robot evaluation frameworks. NVIDIA also introduced the OSMO edge-to-cloud compute framework to simplify robot training workflows across distributed infrastructure.

According to TechCrunch, NVIDIA continues building backbone technology for physical AI, including robots and autonomous vehicles that perceive and interact with the real world. The Alpamayo-R1 model specifically targets autonomous driving research with open reasoning vision language capabilities.

Global partners across robotics, automotive, and industrial sectors are deploying NVIDIA physical AI technologies. Boston Dynamics, Caterpillar, Franka Robotics, Humanoid, LG Electronics, and NEURA Robotics debuted new robots and autonomous machines built on NVIDIA technologies. The partner ecosystem extends to ABB Robotics, AGIBOT, Agility, CMR Surgical, FANUC, Figure, Hexagon Robotics, KUKA, Medtronic, Skild AI, Universal Robots, World Labs, and YASKAWA.

NXP Semiconductors announced collaborative robotics solutions for secure, reliable real-time data processing and transport, integrating NVIDIA humanoid robotics solutions into NXP's safe, secure edge portfolio. NXP stated this collaboration cuts development costs and speeds time to market for physical AI applications.

Industry leaders including Agile Robots, Cadence, and Dassault Systèmes are also building on with these tools. The breadth of adoption signals NVIDIA's platform position across physical AI sectors.

Physical AI enables autonomous systems including cameras, robots, and self-driving cars to perceive, understand, reason, and perform complex actions in the physical world. Previously, autonomous machines could not perceive and sense their surroundings effectively. Physical AI closes this gap, allowing robots to interact with and adapt to real-world environments.

The technology requires tight integration of perception, reasoning, and action in real-time systems. NVIDIA's platform approach addresses this through layered software and hardware integration, from edge sensors through cloud training infrastructure.

Bloomberg interviewed Craig McDonnell, Business Line Managing Director Industries at ABB Robotics, about the partnership with NVIDIA and outlook for physical AI technologies. ABB's involvement highlights industrial manufacturing as a key deployment domain for these capabilities.

NVIDIA is expanding its open model families to power agentic, physical, and healthcare AI, enabling developers and scientists to build intelligent systems that reason and act across digital and real-world environments. The company first advanced this open model development at NeurIPS 2025, releasing tools for speech, safety, and autonomous driving alongside the Alpamayo-R1 model.

New Nemotron, Cosmos, and GR00T models boost multimodal reasoning, physical AI simulation, and biomedical research capabilities. NVIDIA positions open models as essential to advancing innovation at global scale, allowing researchers worldwide to build on shared foundations rather than reinventing core capabilities.

The strategy mirrors NVIDIA's broader approach: provide the computational and software infrastructure that enables an ecosystem, then capture value through hardware and platform services as that ecosystem scales.

Developers can now access NVIDIA's physical AI skills and tools through its open-source repositories, with frameworks designed to reduce complex workflows to agent-executable instructions. The Isaac Lab-Arena evaluation platform and OSMO edge-to-cloud framework specifically target developer productivity in robot training pipelines.

The immediate opportunity lies in integrating these models into specific application domains: warehouse logistics, surgical robotics, autonomous driving, and industrial inspection. Partners are already deploying across these sectors, suggesting reference implementations will emerge quickly.

Competitive pressure on other AI hardware and platform providers will intensify as NVIDIA's ecosystem consolidates. Companies including Intel, AMD, and emerging chip startups must demonstrate comparable software ecosystems to compete for physical AI workloads. Cloud providers Amazon, Google, and Microsoft face strategic decisions about partnering with or competing against NVIDIA's vertically integrated stack.