Nvidia introduced the Cosmos platform and other powerful tools for developing humanoid robots

At CES 2025, Nvidia CEO Jensen Huang presented an updated and expanded set of tools to accelerate the development of humanoid robots. Released back in the spring, Isaac’s GR00T suite for creating synthetic motion sets is now supported by the Cosmos platform – a digital representation of the world around us, created from a colossal amount of diverse data. This will bring closer the day when robots enter human society.

Image source: NVIDIA

The humanoid robot market is expected to reach $38 billion over the next two decades. To meet this significant demand, especially in the industrial sector, Nvidia has introduced not only an extensive set of programming and design tools, but also a collection of basic robot models, data pipelines and frameworks. All this is designed to speed up the development of the next generation of humanoid robots.

Initially, the Isaac GR00T (Generalist Robot 00 Technology) project was focused on the field of simulation training of robots in a wide range of movements. An operator wearing a virtual reality headset, such as Apple Vision Pro (process GR00T-Teleop), performed actions with his hands, and the system broke them down into phases of limb movement and created many alternative movements. This made it possible to significantly speed up learning, eliminating the need to reproduce all human manipulations. Even a small sample of operator actions could result in the creation of an avalanche of synthetic sequences thanks to the computing resources of Nvidia platforms.

After collecting data from the operator’s actions, the GR00T-Mimic process replicates the captured manipulations into a rich synthetic set of movements. The Nvidia Omniverse and Nvidia Cosmos platforms exponentially expand this set by randomizing and scaling actions in 3D space. This data is then used to train robots to move and interact with their environment efficiently and safely in Nvidia Isaac Lab, a modular, open-source platform for robot training.

One of the main announcements of CES 2025 was the Nvidia Cosmos platform, which “bridges the gap between simulation and reality.” These are pre-trained models that reflect the fundamentals of the physical world and are designed to train artificial intelligence taking into account physical processes. The models were trained on 18 quadrillion tokens, including data from 2 million hours of autonomous driving, robotics, drone footage and synthetic sources.

The Cosmos platform not only helps generate large data sets, but also minimizes the gap between simulation and reality by expanding the scale of images from 3D to the real world. The combination of Cosmos with Omniverse, a platform for developing APIs and microservices for building 3D applications, is critical because it minimizes potential errors associated with world models and ensures high accuracy through physically based modeling.

Nvidia has already found interested clients among leaders in the development of humanoid robots, such as Boston Dynamics and Figure. The use of Nvidia Isaac GR00T, Omniverse and Cosmos will significantly advance the robotics industry, bringing closer the day when humanoid robots will become an integral part of human civilization.