Nvidia announced a major set of products for robotics and physical AI at SIGGRAPH. The company introduced new Cosmos world models and upgrades to Omniverse and server hardware. The most notable model is Cosmos Reason, a 7 billion-parameter reasoning vision language model built for physical tasks.
Cosmos Reason is designed to give agents a longer memory and a physics-aware view of the world. Nvidia says the model can act as a planning engine for robots and can help with scene understanding and video analytics. The company describes uses like data curation and step-by-step planning for embodied agents.
How Transfer 2 and synthetic data fit in
Nvidia also announced Cosmos Transfer 2, which helps generate photorealistic synthetic data from 3D scenes. The tool takes spatial inputs like depth maps or segmentation and turns those into varied training images or videos. A distilled version of Transfer makes this work fast enough to run on RTX Pro servers for many developers.
The company added new neural reconstruction libraries that let developers build accurate 3D scenes from sensor streams. Nvidia said it is integrating these tools into popular simulators, including CARLA and into Omniverse, so teams can train agents on realistic virtual copies of the physical world. These tools aim to reduce the cost of collecting and labeling real-world data.
New servers and cloud services for robotics workflows
Nvidia introduced hardware for robotic development, including RTX Pro Blackwell Servers designed for robotics workloads. The company also highlighted its DGX Cloud offering, which manages compute and model workflows at scale. Nvidia positions these products as a single path from model training to simulation and deployment on real machines.
Robotics teams need more than perception models. They need models that can plan and reason about physics and that can produce training data at scale. Nvidia built Cosmos to handle those tasks and to link simulation to real systems. The company sees physical AI as the next growth area beyond data center inference and cloud services.
Early partners and use cases
Nvidia highlighted early adopters in enterprise settings where physical AI powers logistics, security, and customer service agents. The company also named partners that are testing the models in manufacturing and autonomous vehicle workflows. The aim is to let teams test ideas in simulation and then move to real hardware with less manual data work.
Teams that build robots will need to add simulation-centric steps to their pipelines. They will need to validate synthetic data and test agent plans in multiple virtual conditions before any real-world runs. Developers will also need to learn the new Omniverse libraries and test performance on RTX Pro hardware or in DGX Cloud. Nvidia says sample code and SDK updates will help shorten the learning curve.
Simulated data cannot capture every nuance of the real world. Teams must still test and verify models carefully in controlled settings. Models that reason about physical action require safety checks and rule-based guards. Nvidia notes that these tools speed development but do not remove the need for human oversight and robust testing before deployment.
How this fits into Nvidia’s broader strategy
Nvidia is expanding from chips and data center tools into a full stack for physical AI. The company is linking model research with developer tools, simulation, and server hardware. This effort follows earlier Cosmos work and recent industry partnerships that show Nvidia wants to be a central vendor for robotics and autonomous systems. The new releases aim to make the stack easier to use for teams that must scale from experiment to product.
Nvidia has put large pieces in place to help robots learn and act in the physical world. The new Cosmos models add reasoning and synthetic data tools. The Omniverse updates and the server products aim to make those models practical for teams that build robots and industrial agents. The coming months should show how well the stack performs in real applications.