NVIDIA Shows 10 Omniverse Projects Built by Developers Directing Frontier AI Agents

NVIDIA's Into the Omniverse post of 5 November 2026 collects more than ten projects in which developers direct AI agents to build or test simulations on Omniverse libraries. The developers write prompts and review the output. The models named include GPT-6 Astra, which was used most often, Anthropic's Claude, and Cosmos3-Nano for varying weather and lighting in recorded driving video.

The libraries involved are ovrtx for rendering, ovstage for scene updates, ovstream for streaming, ovphysx for physics and ovui for interface, alongside Isaac Sim, Isaac Lab, Newton physics and Warp. Houdini and Blender also appear in the toolchain. NVIDIA does not give version numbers for the libraries or models.

The warehouse stress test is a good concrete result. An agent built a scene of 80 conveyors, 40 robot arms and 800 boxes that ran at 30 frames per second in full HD for a two-minute test, with 640 successful box placements. Doubling the scene reduced performance, and the slowdown was traced to the physics and robot-control pipeline rather than rendering.

Other projects include a kitchen scene with 180 placements from 63 library assets, a simulated Unitree G1 robot clearing a hurdle in 64 of 100 trials, and a gas-turbine tool that tests a stuck fuel valve and an exhaust-sensor error. Agents created two sensor digital twins from scratch and improved two existing ones over about three days.

NVIDIA also points to its SimReady Central Library, announced at GTC Berlin, which offers more than 2,000 simulation-ready USD assets hosted on Hugging Face. Several open-source resources are linked, including Houdini MCP tools.

What this means for 3D work is that scene assembly, the slow placing and checking of assets, is becoming something an agent can do under supervision. The results above also show the honest limit: the agent built the scene, and a human still found that physics was the bottleneck.

We would treat it as a preview of how real-time 3D pipelines are going to be staffed, with artists directing and checking rather than placing every object by hand.