NVIDIA makes the case for open world models in physical AI development

NVIDIA is arguing that open world models will become an important foundation for physical AI, the category of systems that must understand and act in real environments rather than only generate text or images. In a new overview of its approach, the company said such models can learn physical relationships, predict what may happen next and help teams generate data, run simulations and specialize a system for a particular robot, vehicle or vision application. The announcement centers on the company’s Cosmos 3 model family and Omniverse libraries, but it is also a statement about the importance of adaptable models and shared simulation tools in a field where deployments rarely look identical.

Physical-AI teams face a data problem that differs from many software-only applications. A robot or vehicle must cope with changing lighting, weather, objects, sensors and human behavior, including rare events that may be difficult or unsafe to collect in the real world. NVIDIA says world models can generate physically grounded world and action data, simulate future states and provide a starting point for systems that need to reason about those conditions. The company describes that capability as a way to test policies and create more varied training environments before a system is placed into a live setting. It does not remove the need for real-world validation, but it can make early development less dependent on every scenario occurring naturally.

The company’s Cosmos 3 family is positioned as a common base for several related tasks. NVIDIA says it combines vision reasoning, world generation and action prediction, allowing developers to use one family to understand scenes, generate synthetic data, simulate possible states and build specialized world-action models. The lineup includes Cosmos 3 Super, a 64-billion-parameter model for high-fidelity world modeling; Cosmos 3 Nano, a 16-billion-parameter model for more efficient reasoning and post-training; and Cosmos 3 Edge, a four-billion-parameter model for on-device vision reasoning and robot-policy deployment. Those specifications describe NVIDIA’s product family rather than a claim that one model will fit every deployment unchanged.

Openness is central to the company’s argument. NVIDIA says Cosmos world foundation models are available under the Linux Foundation’s OpenMDW 1.1 license, permitting teams to post-train models on their own data and hardware. In physical AI, that flexibility matters because a general-purpose model has not seen every robot design, sensor array, warehouse, road or operating condition a customer may use. NVIDIA presents access to weights, an adaptation-friendly license and post-training tools as practical requirements for closing that gap, not simply as an ideological preference about open models.

The modeling layer is paired with simulation infrastructure. NVIDIA says Omniverse libraries can help developers build simulation-ready worlds, while OpenUSD provides a framework for composing and exchanging complex three-dimensional data across digital twins, simulations and synthetic-data workflows. The intended benefit is a reduction in repeated work when teams change assets, sensor configurations or environmental conditions. In practical terms, a developer may need to alter a camera placement or weather condition without rebuilding every piece of a virtual environment. The announcement presents the tools as building blocks, not as an assurance that simulation alone will produce safe real-world behavior.

NVIDIA also cited use across robotics, autonomous vehicles and vision AI. It named companies working with Cosmos in those areas and pointed to a Cosmos Coalition intended to bring together model builders, developers and physical-AI organizations around models, research and evaluation methods. The company has recently expanded that coalition in Japan, where participating organizations plan to develop open world models for factories, logistics, agriculture, construction, healthcare and transportation. These examples show where NVIDIA expects the approach to be applied, although they do not establish that every industry has reached the same level of deployment or maturity.

The broader significance is that physical AI is moving toward a stack that includes models, specialized training data, simulation environments, evaluation and domain-specific adaptation. NVIDIA’s case for open world models is strongest where a customer needs control over that entire chain rather than a fixed API response. The open questions concern cost, quality and transfer to reality: how accurately can a simulated world represent an uncommon event, how much post-training is needed for a particular deployment and how will teams measure safety before an automated system acts around people? The company’s August update offers a roadmap for those problems, but the proof will depend on the outcomes of the systems built on top of it.