NVIDIA released Cosmos 3, an open foundation model that combines world simulation, action generation and reasoning to train robots on real-world tasks.
On May 31, 2026 NVIDIA announced Cosmos 3, an open world foundation model for Physical AI. It is the company’s first fully open omnimodel, able to understand and generate several data types at once.
Its mixture-of-transformers architecture pairs a reasoning transformer with an expert generation transformer. This lets the model reason about physical interactions before predicting what happens next.
According to NVIDIA it was trained on one of the largest multimodal physical-AI datasets, with billions of samples across text, image, video, sound and action trajectories. It works as a vision-language model, a world model, a video foundation model and a backbone for teaching robots specific tasks.
It comes in Super and Nano variants, with Edge to follow, alongside a Cosmos Coalition whose partners include Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI. It is the direction Ergonet is building toward: machine vision and world models that let robots and automation systems understand a space before acting in it.
Source: NVIDIA