NVIDIA Unveils Cosmos: End-to-End Physical AI Stack, Omniverse Upgrades and Blackwell Servers

NVIDIA announced a broad new stack at SIGGRAPH 2025 focused on accelerating physical AI for robotics, autonomous vehicles, and industrial automation. The release bundles reasoning world models, synthetic data generators, simulation and rendering libraries, and infrastructure tuned for embodied intelligence workflows.

Cosmos Reason: reasoning for embodied agents

At the center of the launch is Cosmos Reason, a 7-billion-parameter vision-language reasoning model built for robots and other embodied agents. Cosmos Reason combines memory for spatial and temporal contexts with an understanding of physical dynamics, enabling step-by-step planning in complex, real-world scenarios. The model consumes structured video and sensor inputs such as segmentation maps and LIDAR, then produces both high-level instruction parsing and low-level action suggestions to guide navigation and manipulation.

Key capabilities highlighted by NVIDIA include:

• Memory and physics awareness that help models reason about object permanence, forces, and temporal sequences.
• A planning engine that maps visual and sensor streams into ordered actions, useful for robot planning, data curation, and video analytics.

Cosmos Transfer: faster synthetic dataset generation

Cosmos Transfer-2 focuses on converting 3D scenes and spatial control inputs into large-scale synthetic datasets for training and validating robot policies. This pipeline reduces time and cost for producing realistic training data and makes it easier to cover edge cases like unusual lighting, occlusions, and diverse weather conditions.

NVIDIA also announced a distilled transfer variant optimized for speed so developers can iterate quickly on dataset creation. The broader Cosmos WFM family spans Nano, Super, and Ultra tiers (roughly 4 billion to 14 billion parameters) and supports fine-tuning for different latency and fidelity trade-offs, from real-time streaming to photorealistic rendering.

Simulation and rendering libraries in Omniverse

Omniverse receives several upgrades aimed at building lifelike virtual worlds for training and testing:

• Neural reconstruction libraries to convert sensor captures into high-fidelity 3D scenes using neural rendering techniques.
• Improved integration tools for OpenUSD and the CARLA simulator to standardize conversion and rendering workflows across robotics frameworks such as Mujoco and USD-based pipelines.
• SimReady materials library with thousands of substrate materials for creating realistic virtual environments.
• Isaac Sim 5.0, which brings enhanced actuator models, expanded Python and ROS support, and new neural rendering features for higher-quality synthetic data.

These tools make it easier to build, simulate, and render environments that mirror real-world complexity so policy models can be trained on more representative scenarios.

Infrastructure for robotics workflows

NVIDIA introduced hardware and cloud services to support end-to-end development:

• RTX Pro Blackwell servers designed specifically for simulation, training, and inference workloads in robotics development.
• DGX Cloud for scalable cloud-based management of physical AI workflows, enabling distributed teams to develop, train, and deploy agents remotely.

This unified architecture targets the full lifecycle of an embodied AI project, from dataset generation and simulation to model training and deployment.

Industry adoption and licensing

Major robotics and mobility companies including Amazon Devices, Agility Robotics, Figure AI, Uber, and Boston Dynamics are piloting Cosmos models and Omniverse tools. Use cases include generating training data, building digital twins, and accelerating deployment in manufacturing, transportation, and logistics.

NVIDIA is making Cosmos models available via APIs and developer catalogs under a permissive license to support both research and commercial projects.

What this means for physical AI

NVIDIA is positioning physical AI as a full-stack challenge: smarter reasoning models, richer simulation tooling, and scalable compute must come together to shrink the gap between virtual training and real-world performance. By combining Cosmos world models, Omniverse simulation libraries, and Blackwell-powered infrastructure, the company aims to reduce costly trial-and-error and enable more robust autonomy for robots and embodied agents.