AI Finally Understands Real-World Physics: Nvidia Cosmos Paves the Way for Intelligent Robotics

“The ChatGPT moment for general robotics is just around the corner,” CEO Jensen Huang said at CES 2025.

Jensen Huang
No comments Twitter Flipboard E-mail
enrique-perez

Enrique Pérez

Writer

Consumer tech and information society editor. Despite my studies in Physics, I've been writing about tech, image and sound, digital economy, legislation and data protection for over ten years. Interested in projects that aim to improve society and democratize access to technology. LinkedIn

Nvidia has started the year off on an impressive note. At CES 2025, CEO Jensen Huang unveiled the company’s new products, including the new RTX 50 graphics card. More importantly, he showcased various initiatives centered around the technology that has propelled Nvidia to the forefront: artificial intelligence.

One exciting development is Nvidia Cosmos, the first AI platform capable of “understanding” real-world physics.

The Cosmos models are open source and come in three sizes: Nano, Super, and Ultra. They’re available in Nvidia’s catalog and on Hugging Face. According to Nvidia, developing these models is quite costly. They’re trained on millions of real-world data points, including 20 million hours of video. This enables the AI models to “see” how the world operates.

Huang claims Cosmos will dramatically accelerate the training time for intelligent robots and advanced autonomous vehicles. “The ChatGPT moment for general robotics is just around the corner,” he said during Nvidia’s keynote.

Several leading companies in the fields of robotics and autonomous vehicles will soon begin testing Cosmos. These include 1X, Agile Robots, Agility, Figure AI, Foretellix, Fourier, Galbot, Hillbot, IntBot, Neura Robotics, Skild AI, Uber, Virtual Incision, Waabi, and XPENG.

Brand logos

Cosmos models offer interesting capabilities, particularly in the realm of the so-called “physical AI.” One notable application is the ability to analyze scenarios in videos. For instance, you can instruct the AI model to “find a snowy road” or to “search for a warehouse with limited space.”

This concept also enables the generation of videos that adhere to the laws of physics. One of the challenges faced by many generative AIs has been failing to produce physically accurate movements. The Cosmos model aims to address this issue.

Nvidia Cosmos

This is merely the beginning. Like all AI models, the results and use cases will lead to a reevaluation of the original model. As Nvidia Cosmos is used, it’ll gain a better understanding of how different scenarios operate.

Cosmos is specifically designed for robots and autonomous vehicles. Overall, it helps machines comprehend their surroundings.

Image | Nvidia [1, 2]

Related | Nvidia Has Broken Its Revenue Record Once Again. Yet, the Company Is Much Smaller Than It May Seem

Home o Index