Nvidia has released Cosmos 3, an open-source foundation model trained on a diverse dataset of physical world interactions. The model learns to predict how objects move, respond to forces, and interact with one another—core capabilities for robotics, autonomous systems, and simulation.
The release is significant because foundation models for physical reasoning are rare. Most large models are trained on text and images; Cosmos is explicitly built to understand physics. A robot using Cosmos can reason about grasp points, force application, and collision prediction without explicit programming.
Nvidia positioned Cosmos as open-source, meaning researchers and companies can fine-tune it for their own robotics and simulation tasks. This democratizes physical AI—the field that powers autonomous systems, warehouse robots, and industrial automation.
The timing matters. Robotics is entering a phase where embodied AI (AI trained on real-world physical interaction) is becoming competitive with simulation-only approaches. Cosmos 3 accelerates that shift by providing a pre-trained physical reasoning backbone that reduces the data and compute required to build capable robots.
