Nvidia has introduced a new suite of AI tools to advance robotics and other physical AI applications.
The announcement came at the SIGGRAPH conference on Monday, highlighting a stronger push into robotics beyond the company’s data center market.
Cosmos Reason
The core of this release is Cosmos Reason, a vision–language model with 7 billion parameters.
It is designed to help robots and AI agents “reason” about their environment. This includes understanding physics, recalling past interactions, and planning the next steps.
Nvidia says the model can support data curation, robotic planning, and video analytics.
Additional Cosmos Models
Nvidia also released Cosmos Transfer-2. This model can rapidly generate synthetic datasets from 3D simulations or spatial control inputs.
It is joined by a distilled, faster version of Cosmos Transfer for speed-sensitive workflows.
Together, these models can create training data for robots without requiring extensive real-world testing.
Advanced 3D Reconstruction
The company introduced new neural reconstruction libraries, known as Omniverse NuRec.
These include a rendering method called 3D Gaussian splatting, which reconstructs realistic 3D environments from sensor data.
This rendering capability will also be integrated into CARLA, an open-source simulator widely used by robotics developers.
Hardware And Cloud Support
To handle these demanding workloads, Nvidia unveiled the RTX PRO Blackwell Server to provide a unified architecture for robotics development.
Additionally, the DGX Cloud platform provides cloud-based infrastructure for scaling projects without the need for on-premises hardware.
Early Industry Adoption
Several companies are already exploring the technology. Amazon Devices & Services, Boston Dynamics, Figure AI, Hexagon, Uber, and Zoom are among the first to integrate these tools into their development processes.
