Nvidia’s new world model helps robots navigate the world Nvidia has introduced Cosmos 3, an open AI world model trained on extensive multimodal data to enhance the understanding and prediction capabilities of physical systems like robots and autonomous vehicles. This model focuses on generating action data, enabling machines to navigate and manipulate the physical world, and is positioned to become a foundational platform for physical AI development. Nvidia aims for Cosmos 3 to facilitate the creation of more generalizable physical agents by allowing them to understand and plan actions within real-world environments.

• Nvidia launched Cosmos 3, an open AI world model for physical systems.
• The model helps robots and autonomous vehicles understand and predict real-world environments.
• Cosmos 3 is trained on 20 trillion tokens of multimodal data, focusing on action data for machine movement.
• Developers can use Cosmos 3 to simulate actions and build task-specific models.
• It generates action data like robot joint angles and trajectories to train machines.
• Cosmos is open, allowing hardware makers to customize it for their needs.
• Nvidia is forming a coalition of supporting companies.
• The model can generate rare or dangerous scenarios for training.
• Two versions, ‘super’ and ‘nano’, are immediately available, with an ‘edge’ version coming soon.
• World models are a growing AI area for enabling real-world tasks.
• Nvidia’s strategy is to be the starting point for AI development in the physical world. Continue reading https://www.axios.com/2026/06/01/nvidia-ai-push-cosmos-3-world-model