NVIDIA Cosmos 3: Models That Reason and Act in the Physical World of Robots

NVIDIA has introduced Cosmos 3 on the NVIDIA Developer Blog, the new generation of its Cosmos platform: a family of models for so-called physical AI. These models are meant to understand how the real world works, predict how events in it will unfold, and generate actions for robots, autonomous vehicles and "smart" spaces.

What is physical AI, and why does it need such models

Unlike language models, which work with text, or computer vision models, which simply recognize objects in an image, physical AI has to solve a harder problem: understanding the surrounding world as it changes. NVIDIA frames this as a three-step process. The system first has to work out what is happening in its physical environment right now, then predict what is likely to happen next, and only after that generate the right action or plan of action. Cosmos targets exactly this "perception → prediction → action" pipeline.

Key facts about the platform:

• Developer: NVIDIA
• Generation name: Cosmos 3
• Task category: "world foundation models" for physical AI
• Target devices: robots, autonomous vehicles, "smart" spaces
• Source of the announcement: NVIDIA Developer Blog

How Cosmos fits into the robotics ecosystem

To work safely and usefully, robots and self-driving cars need more than the ability to "see" their surroundings. They need a world model that lets them anticipate the consequences of their own actions, and those of other agents in the environment, before any physical movement has happened. Such models used to be built individually for specific hardware and scenarios, which made them expensive and hard to transfer between different types of robots.

A general-purpose platform like Cosmos instead provides a base layer of "physics understanding" of the world. Developers of specific robotic systems, from industrial manipulators to service robots, can build their own, more specialized models on top of it without starting training from scratch each time. This approach, a general-purpose model as the foundation on which narrow solutions are then fine-tuned, has already proved effective for language and vision models, and NVIDIA appears to be betting that the same logic will work for physical AI.

Which areas the technology will affect

The practical significance of such models reaches well beyond robotics alone. Autonomous transport is the second most important category: there, the system must accurately predict the behavior of other road users a fraction of a second ahead in order to make safe decisions in real time. The third area is "smart" spaces: factory floors, warehouses and other environments equipped with sensors and cameras, where physical AI can coordinate many devices at once, relying on a shared model of what is happening in the room.

The arrival of such a platform from NVIDIA, one of the main suppliers of computing infrastructure for AI overall, signals that in 2026 the industry's focus is gradually shifting from models that work exclusively with text and images to models that are expected to act in the physical world and bear responsibility for the consequences of their decisions.

For developers of specific robotic systems, the practical value of such a platform lies in shortening the time between an idea and a working prototype. Instead of spending years collecting and labeling their own dataset of robot interactions with the physical environment, a team can start from an already trained "world foundation" model and adapt it to a specific task, such as a particular type of manipulator or a warehouse logistics scenario. This is the path computer vision and natural language processing took a few years ago, when the shift from training models from scratch to fine-tuning ready-made base models sped up the delivery of applied products by an order of magnitude.

NVIDIA clearly expects to repeat that effect for robots and autonomous systems.