NVIDIA presented examples of agents using Omniverse tools to create simulations for warehouse robots, autonomous driving, sensor validation and other tasks.
On October 8, 2026, NVIDIA detailed a collection of developer projects in which AI agents helped assemble simulation applications using NVIDIA Omniverse libraries. The examples show agents responding to natural-language direction while developers review outputs and steer revisions; the libraries supply GPU-accelerated physics, rendering and sensor simulation.
The demonstrations cover several stages of robotics and autonomous-vehicle development. In one, NVIDIA’s Frank DeLise used GPT-6 Astra to connect warehouse assets, robot models, physics and interface components into an interactive humanoid simulator. Another project used Astra to build a reusable San Francisco Market Street environment for autonomous-driving tests, linking traffic, assets and RTX sensor simulation so developers could compare driving models and examine how scene changes affect results.
NVIDIA also described an effort to compare simulated camera and raw LiDAR outputs against recorded data. Over roughly three days, Ashley Reid guided Astra and Claude Fable 5 agents through creating two digital twins and improving two existing ones. The workflow used measured differences to identify issues such as missing objects, geometry and materials.
Other examples applied simulation to robot skills and product disassembly. In an experimental sports project, a simulated Unitree G1 cleared a single hurdle in 64 of 100 trials. For a car-suspension disassembly task, an agent helped model the part and design a wrench that could reach its bolts; NVIDIA reported successful simulated removal of a suspension component.
The projects are demonstrations of agent-assisted workflows, not evidence that agents can independently validate real-world systems. Across the examples, developers remained involved in directing work and assessing results. NVIDIA also showed agents assembling a browser-based International Space Station model with telemetry and converting captured room geometry into an editable environment for interaction testing.