NVIDIA Shows AI Agents Building Omniverse Simulations, From Gas Turbines to Hurdling Humanoids
Oct 30, 2026 — Building a physics simulation used to mean weeks of wiring together assets, solvers and renderers by hand. NVIDIA now says engineers on its own teams are handing much of that work to frontier AI agents and steering them in plain English. In an "Into the Omniverse" post published Thursday, Oct. 29, the company showcased more than a dozen projects where developers directed models such as OpenAI's GPT-6 Astra to assemble simulation apps using NVIDIA Omniverse libraries.
The pitch: developers direct, agents build
According to NVIDIA, the workflow is simple to describe: developers give AI agents natural-language instructions, review the results and guide changes, while Omniverse supplies GPU-accelerated physics, rendering and sensor simulation. NVIDIA says it will keep adding examples from its teams and the broader ecosystem.
Highlights from the showcase
Warehouse stress test. NVIDIA principal engineer Bhavik Sanganee used Astra to build and scale a warehouse simulation. According to the post, with 80 conveyors, 40 robot arms and 800 boxes, the scene held 30 frames per second at full HD during a two-minute test, and the arms completed 640 successful box placements. Doubling the scene again cut performance, and testing traced the slowdown to the physics and robot-control pipeline.
Gas-turbine "what-if" app. Pradeep Kaushik of NVIDIA's Omniverse developer relations team directed Astra to build a digital-twin app around a simulated gas turbine and generator. Users can adjust load, ambient temperature and fuel composition, and NVIDIA says six test procedures cover scenarios including a stuck fuel valve and an exhaust-sensor error. The agent used Cantera for thermochemistry and SciPy for equipment dynamics.
Robo Olympics. Tae Kim, who leads Omniverse engineering and product, used sports videos and text instructions to have Astra build controllers for simulated Unitree G1 humanoids. In one experiment, according to NVIDIA, the robot cleared a single hurdle in 64 of 100 simulation trials.
Cable routing for electronics assembly. Robotics marketing manager Oyindamola Omotuyi directed Astra to build a simulation with Isaac Lab, Newton physics and a Robotiq gripper handling a flexible cable for USB-C insertion, including a reinforcement learning component.
Autonomous-driving test bed. Doyub Kim had Astra build "Zero to Alpamayo," a reusable driving-simulation environment based on San Francisco's Market Street, connecting traffic, RTX sensor simulation and the Alpamayo driving model.
Sensor-matched digital twins. Ashley Reid, who works on RTX sensor validation, directed Astra and Claude Fable 5 agents to compare simulated camera and lidar output with recorded data. NVIDIA says the agents created two digital twins from scratch and improved two more over about three days.
Other examples include reconstructing an empty building model from site scans, converting an ABB YuMi robot CAD file into a simulation-ready asset, assembling a kitchen scene from the SimReady Central Library, a humanoid warehouse simulator, a robotic car-suspension disassembly test, and a browser-based International Space Station model built from NASA assets.
The caveats NVIDIA included
The post is careful about what these demos prove. For the YuMi robot, NVIDIA notes the tests showed behavior under simulated conditions and that estimated physical properties "were not calibrated against the real robot." For the cable-routing work, next steps include using sensor information a physical robot could actually obtain. These are prototypes and internal experiments, not shipping products.
Why it matters for AI and games
Omniverse is aimed at industrial and robotics work, but the underlying pattern, using an agent to stitch together assets, physics and rendering through conversation, overlaps heavily with how game developers prototype levels and systems. NVIDIA is also promoting "agent skills" and libraries such as ovrtx (rendering), ovstage (scene updates) and ovstream (browser streaming) as the building blocks agents call on.
What's confirmed / not confirmed
Confirmed (NVIDIA blog, Oct. 29, 2026): - NVIDIA staff used frontier AI agents, including GPT-6 Astra, with Omniverse libraries to build the showcased simulation projects. - The warehouse demo figures (80 conveyors, 40 arms, 800 boxes, 30 fps at full HD, 640 placements) and the 64-of-100 hurdle result are NVIDIA's own reported numbers. - SimReady Central Library offers more than 2,000 simulation-ready USD assets, per NVIDIA.
Not confirmed: - Independent replication of any of the demo results. - How much human correction each project required beyond NVIDIA's descriptions. - Any commercial release dates for the prototypes shown; NVIDIA presents them as examples, not products.
Video
Video by NVIDIA on YouTube.
Sources
- Official: NVIDIA Blog, "Into the Omniverse: How Developers Turn Ideas Into Simulations With Frontier AI Agents" (Oct. 29, 2026): https://blogs.nvidia.com/blog/developers-simulation-frontier-ai-agents/
Up: AI