THE CRUNCH
NVIDIA has published a set of demonstration projects in which its staff direct frontier AI agents, chiefly its GPT-6 Astra model, to build working simulation applications on the Omniverse platform. The idea is that turning a simulation concept into a running application normally means assembling assets, wiring up physics and rendering, and checking the scene behaves as intended. In these examples, developers give the agents natural-language instructions, review what comes back and steer the changes, while Omniverse libraries supply GPU-accelerated physics, rendering and sensor simulation.
The projects span several domains. One turned a SimReady warehouse and humanoid robot into an interactive simulator with first- and third-person views, with the agent connecting the physics, scene, rendering and interface libraries. Another, built on a recreation of San Francisco's Market Street, gave autonomous-driving teams a test bed for tracing how scene or sensor changes affect driving behaviour, with a separate Cosmos3-Nano experiment varying weather and lighting to compare responses to the same scenario.
Sensor fidelity got its own treatment: agents guided by RTX sensor validation work compared simulated camera and LiDAR outputs against recorded data, creating two digital twins from scratch and improving two existing ones over roughly three days, with acceptance hinging on camera and LiDAR metrics. Elsewhere, an agent designed a wrench a simulated robot could use to remove a car suspension bolt, and an engineering director assembled NASA assets into a browser-ready International Space Station model with telemetry from a single prompt.
The results are promising but not uniformly so. In the Robo Olympics experiment, where simulated Unitree G1 humanoids performed sports movements, the robot cleared a single hurdle in 64 of 100 simulation trials, with the trials feeding back into timing and control improvements. These are experiments by NVIDIA's own teams rather than shipping products, so they show a workflow in development rather than a finished capability.