At SIGGRAPH 2025 in Los Angeles, Nvidia introduced a major expansion of its Cosmos world models, simulation libraries, and hardware tailored for physical AI. Demonstrations on the show floor illustrated how these additions could cut development time for robotics, autonomous shuttles, and factory automation. Attendees from research labs and industrial teams gathered to explore methods for bringing virtual training closer to real-world operation across tasks like material handling, quality inspection, and site mapping.
The centerpiece is Cosmos Reason, a 7-billion-parameter vision-language model designed for robots and embedded agents. It combines advanced memory for spatial and temporal context with an understanding of gravity, friction, and collisions to plan multi-step workflows. By streaming segmentation maps, LIDAR point clouds, and video frames into its reasoning core, the model interprets high-level instructions—such as “sort these items” or “inspect that crate”—and generates precise actuator commands.
To streamline training data creation, Nvidia rolled out Cosmos Transfer-2. Engineers can feed 3D scene descriptions or spatial control signals into the system, then obtain synthetic images under varied lighting, weather, and material setups. This approach cuts provisioning time by over half for reinforcement learning and policy testing. A lighter Distilled Transfer variant offers faster turnaround for quick iterations, letting teams cycle through hundreds of customized data batches daily with minimal manual oversight.
The Cosmos WFM product line spans three sizes—Nano (4B parameters), Super (9B), and Ultra (14B). Each model targets specific latency and fidelity requirements. Nano suits low-bandwidth devices for real-time camera streaming. Super balances performance for mixed simulation and inference tasks. Ultra focuses on high-resolution rendering for design validation or virtual reality previews. All versions support fine-tuning on domain-specific datasets to meet diverse needs across robotics, AR, and industrial inspection applications.
Nvidia’s Omniverse environment received new neural reconstruction libraries that convert sensor logs from multi-camera rigs and depth scanners into detailed 3D scenes. Integration tools bridge OpenUSD pipelines with the CARLA simulator, simplifying workflows between Mujoco-based control modules and Nvidia’s USD architecture. A SimReady Materials Library offers over ten thousand ready-made substrates—metals, plastics, textiles—for constructing realistic virtual facilities and digital twins for robotics tests.
In Isaac Sim 5.0.0, actuator models reflect real-world torque curves and expanded Python and ROS2 APIs speed scenario scripting. Neural rendering modules simulate camera noise, motion blur, and depth artifacts for more robust training. On the hardware front, RTX Pro Blackwell servers link GPUs via NVSwitch, provide up to 256 GB per GPU, and exceed 100 TFLOPS of mixed-precision compute. DGX Cloud runs large-scale physical AI workflows on shared clusters.
Several organizations have launched pilot programs using these tools. Amazon Devices simulates last-mile sorting in digital replicas of fulfillment centers. Agility Robotics tests biped balance and obstacle courses entirely in simulation before hardware trials. Uber prototypes sidewalk rover navigation with detailed traffic and pedestrian models. Boston Dynamics examines quadruped stability across varied terrains. These testbeds produce annotated datasets, custom digital twins of facilities, and continuous performance benchmarks for refinement.
All Cosmos offerings are available through Nvidia’s developer catalog and APIs under a permissive license for research and commercial applications. Pretrained weights can be fine-tuned on custom environments and deployed on local RTX systems or in DGX Cloud clusters, leveraging Nvidia’s global data center footprint and flexible support options. Nvidia presents this suite of Cosmos models, Omniverse libraries, and Blackwell-powered servers as a unified solution that cuts trial cycles, lowers field errors, and speeds autonomous machine deployment.