NVIDIA Omniverse: Simulation Examples Featuring Astra Agents Shared

In a blog post published on 9 October 2026, NVIDIA introduced simulation projects developed with AI agents by developers using Omniverse libraries. The examples range from warehouse robots and autonomous vehicle testing to digital twins and browser-based 3D applications. In the workflows described by NVIDIA, developers give agents natural-language instructions, review the results, and guide them through changes.

How was Astra used in the warehouse simulator?

NVIDIA Omniverse product manager Frank DeLise brought together a SimReady warehouse environment and a humanoid robot in an interactive simulator. In this custom example, Astra was used to prepare application code connecting ovphysx for physics, ovstage for scene updates, ovrtx for rendering, and ovui for the user interface. SimReady also helped set up the physical scene in the simulation. The resulting environment lets users inspect warehouse tasks from first- and third-person viewpoints.

In an autonomous driving project, NVIDIA team member Doyub Kim created a reusable test environment based on Market Street in San Francisco. Astra gradually brought together asset generation, traffic, Omniverse RTX sensor simulation, and Alpamayo driving components. The prototype was used to compare how changes to the scene or sensors affected driving behavior. In a separate experiment, Cosmos3-Nano was used to change weather and lighting conditions in recorded simulation videos.

How were digital twins and robot movements tested?

In work by Ashley Reid of the RTX sensor validation team, Astra and Claude Fable 5 agents evaluated differences between camera and raw LiDAR outputs generated with ovrtx and recorded data. During an iterative process lasting around three days, the agents created OpenUSD scenes or edited existing ones. Changes involving missing objects, geometry, and materials were checked against camera and LiDAR measurements.

In an experiment called Robo Olympics, Astra developed control systems for simulated Unitree G1 humanoid robots using sports videos and natural-language instructions. In one of the trials shared by NVIDIA, the robot cleared a single obstacle in 64 out of 100 simulation attempts. The work used Newton Physics Engine for physics simulation, NVIDIA Warp for compute acceleration, and ovrtx for rendering. The result is presented as an experimental example of evaluating robot movements in simulation.

The blog also describes a browser-based application featuring an OpenUSD model of the International Space Station with telemetry, for which assets were prepared in Blender. Another project created an editable OpenUSD studio scene from stereo camera recordings; door and drawer collision and contact behavior were reviewed using Isaac Sim tests.

What does this mean for architecture and archviz workflows?

This announcement is not a ready-made product specifically intended for architectural design or archviz. Still, creating an editable interior scene from camera recordings and testing door and drawer interactions in simulation offers an approach that architecture and interior design teams examining spatial behavior may want to follow. The source does not say whether these examples are compatible with architectural production software or constitute a validated archviz workflow.

When assessing their suitability for offices and visualization teams, it is worth considering the transfer of OpenUSD scenes, the GPU-accelerated components required, and the suitability of existing assets. NVIDIA’s post does not disclose hardware requirements, licensing terms, pricing, or a general availability timeline. The examples should therefore be viewed as developer projects, not as a complete production solution.