NVIDIA launches the Cosmos world foundation model platform at CES

Background

Robots and autonomous vehicles need large amounts of data about how the physical world looks and behaves. Collecting that data in the real world is slow and expensive, so developers have looked for ways to generate it. NVIDIA positioned Cosmos as a platform for this work, aimed at what it called physical AI.^[1] Synthetic data, which is generated by computer rather than recorded in the world, can cover situations that are rare or dangerous to record, such as unusual road conditions or failures a robot seldom meets.

What happened

On 6 January 2025, at the CES trade show, NVIDIA launched Cosmos, a platform of world foundation models for physical AI. The models generate physics-aware video, which NVIDIA said could be used to train robots and autonomous vehicles.^[1] NVIDIA released the models under its open model license, so they could be downloaded and used by others.^[1]

Key details

Robotics and automotive companies were named as early adopters. The list included 1X, Agility, Figure AI, Uber, Waabi and XPENG.^[1]

The launch was not only the models. It also included a tokenizer, which converts video into a compact form for the models to process, along with guardrails, which are safety checks, and an accelerated video-data pipeline for preparing training material.^[1]

Later in the year NVIDIA presented Cosmos with its Isaac GR00T N1 humanoid model at GTC 2025, the company's major developer conference.^[2]

Impact

The launch placed NVIDIA's physical AI tools in front of robotics and vehicle developers as a set of open models rather than a closed service. The named early adopters show the companies that were expected to use it first. The combination of a data pipeline and open models was meant to make it easier for developers to produce the synthetic training data their systems need.^[1] The later presentation with a humanoid robot model at GTC 2025 tied the platform to NVIDIA's other robotics work.^[2]