At the recent SIGGRAPH 2025 conference, NVIDIA unveiled its Cosmos world model series for robotics developers, with the 7-billion-parameter Cosmos Reason being a standout. Designed specifically for physical AI, this visual language model leverages memory and understanding of physical principles to empower robots and AI agents with advanced reasoning capabilities, enabling them to predict the actions of embodied agents and apply them to data curation, robotic planning, and video analysis.
Additionally, the new Cosmos Transfer-2 model in the series further enhances the efficiency of synthetic data generation, with a streamlined version offering even faster performance in 3D simulations. NVIDIA emphasizes that these models can batch generate synthetic text, image, and video datasets required for training, significantly lowering the development barrier. Furthermore, the accompanying Neural Reconstruction Library integrates novel rendering technology to transform sensor data into 3D realistic simulations and is integrated into the open-source simulator CARLA. The Omniverse SDK has also been updated to provide a more comprehensive toolchain for developers.
To optimize robotics workflows, NVIDIA simultaneously launched two servers: the RTX Pro Blackwell Server, which provides a unified architecture to support development workloads, and the DGX Cloud, which simplifies cloud management. These moves signal NVIDIA's expansion of AI GPU applications beyond the data center into robotics, accelerating the industrialization of physical AI through a full-stack approach encompassing models, tools, and hardware.
