NEW

Nvidia releases smaller world model for robots

Cosmos 3 Edge is an open AI model that connects visual impressions with possible actions. It is built to run near robots, cameras and industrial machines on Nvidia hardware.

Publicerad 21 July 2026, 11.04

AI-generated illustration of a robotic arm and a mobile inspection robot connecting to a local computing device and a visual model of the warehouse environment.

Nvidia has released the Cosmos 3 Edge on the Hugging Face model platform. It is an open AI model with four billion parameters that is built for robots and other systems that need to interpret camera images and act in a physical environment.

A world model tries to describe how an environment changes over time. For a robot, this could mean recognizing an object, judging where it is and anticipating what will happen if the gripper arm is moved. Cosmos 3 Edge can both reason about a scene and create suggestions for actions.

The model is made for local driving near the sensors, also called edge computing. Nvidia states support for RTX Pro, DGX, GeForce RTX and Jetson, among others. The fact that the model can be downloaded does not mean that it works directly in every robot. Developers need to adapt it to the machine, the task and their own data.

Nvidia has also published a version for tasks where a robot picks and places objects, along with scripts for further training. The company describes real-time control on the Jetson Thor, but results on speed, accuracy and top positions in tests come from Nvidia and are not independently confirmed.

Cosmos 3 Edge does not automatically send data to a data center when running locally. It can reduce latency and give the business more control over data flow. Integrity and security still depend on the entire system, including cameras, storage, networking, and how the model has been adapted.

Därför spelar det roll

Smaller models make it possible to move more AI from cloud services to robots, cameras and industrial machines. For Swedish businesses, it can provide faster reactions and less dependence on a network connection. At the same time, it makes testing, physical security and clear accountability more important when an AI system affects the real environment.

Det här kan du göra

  1. Start with a confined test environment where an error cannot harm people, equipment or production.
  2. Measure the model on your own camera, robot and task instead of relying on the vendor's test results.
  3. Keep a physical stop function and clear boundaries for what actions the model may suggest or perform.