Nvidia unveiled its latest artificial intelligence platform, Cosmos, during the Consumer Electronics Show (CES) 2025 on Monday. This innovative platform incorporates a variety of generative world foundation models (WFMs) along with advanced tokenisers, an accelerated video processing pipeline, and protective measures that facilitate the development of physical AI systems such as autonomous vehicles and robots. In a significant move, Nvidia has also open-sourced the WFMs for academic and research use. Furthermore, the company introduced the Llama Nemotron family of AI models at the event.
Nvidia Cosmos World Foundation Models
In an announcement posted on its newsroom, Nvidia detailed the functionalities of the new Cosmos platform, which is designed exclusively to host WFMs alongside various elements that support the training and development of physical AI systems. Physical AI systems refer to machines equipped with mechanical components that can interact with and execute tasks in the real world.
The company acknowledged the substantial financial investment required for training and developing physical AI systems, such as robots and autonomous vehicles, largely due to the need for extensive real-world data and a wide array of testing environments. The WFMs integrated within the Cosmos platform address these challenges effectively.
Nvidia asserted that these world AI models have the capability to generate vast quantities of photorealistic and physics-based synthetic data, which can be utilized in training physical AI systems. Additionally, this data can be employed to evaluate existing robots through exposure to various testing environments. The Cosmos platform further enables developers to create customized models by fine-tuning the WFMs.
Equipped with video search and comprehension features, the AI models help developers locate specific training videos from an extensive database. They can leverage the Nvidia Omniverse platform to create controlled 3D scenarios based on physics. The platform also provides simulation-based training for physical AI applications.
The AI models are accessible under an open model license, allowing developers to preview them through Nvidia’s application programming interface (API) catalogue or from Hugging Face. Nvidia reported that numerous companies specializing in robotics and physical AI have already begun utilizing the Cosmos platform, including 1X, Agile Robots, Agility, Figure AI, Foretellix, Fourier, Galbot, Hillbot, IntBot, Neura Robotics, Skild AI, Virtual Incision, Waabi, XPENG, and Uber.
