NVIDIA expands Cosmos to accelerate physical AI development
The updated Cosmos foundation models expand NVIDIA Omniverse capabilities, improving physical AI training through realistic, scalable synthetic data generation.
The US tech giant, NVIDIA, is expanding its Omniverse ecosystem with new Cosmos open world foundation models designed to accelerate the training of physical AI systems used in robotics, autonomous vehicles and other intelligent machines.
An updated Cosmos Predict 2.5 that integrates Text2World, Image2World and Video2World into a single framework capable of producing consistent, multi-camera synthetic video worlds from text, image or video prompts.
By generating realistic, controllable environments, Cosmos enables developers to create synthetic datasets essential for safely testing AI in dynamic physical conditions that are difficult or hazardous to replicate in reality.
Cosmos Transfer 2.5 adds high-fidelity world-to-world style transfer, letting developers modify terrain, lighting and weather conditions across simulations while improving speed and accuracy.
The foundation models integrate with NVIDIA Isaac Sim and Omniverse libraries, providing a unified pipeline for synthetic data generation, simulation, and testing.
Companies such as Skild AI, Serve Robotics and Zipline use these tools to train robot systems, validate autonomous navigation, and enhance operational safety. Synthetic data reduces costs and accelerates development by bridging the gap between simulation and real-world performance.
Beyond robotics, the ecosystem supports industries such as logistics, mining and energy. Developers and researchers can use OpenUSD-based workflows, Omniverse libraries and NVIDIA Brev to create realistic training data, advancing physical AI that perceives and interacts with the real world safely and efficiently.
Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!
