NVIDIA published a showcase of simulation projects on October 8, 2026, in which developers directed frontier AI agents—including GPT-6 Astra—to build applications with its Omniverse libraries, according to a company blog post.
The demonstration illustrates how large language models can automate the assembly of physics, rendering and sensor-simulation pipelines, potentially shrinking the manual engineering effort required to create digital twins and robotics testbeds. The announcement includes no independent verification of the agents’ performance, pricing or error-rate data.
Omniverse libraries provide GPU-accelerated physics (ovphysx), scene management (ovstage), rendering (ovrtx) and user-interface tools. In each project, developers described a simulation goal in natural language and iterated with the AI agent, which selected and connected the relevant Omniverse components.
Building simulations with natural-language instructions
Frank DeLise, an NVIDIA product manager, used GPT-6 Astra to combine ovphysx, ovstage, ovrtx, ovui and SimReady foundation libraries with a warehouse asset, producing an interactive humanoid simulator with first- and third-person views, the blog post states.
Doyub Kim, a simulation-technology manager, had Astra build a reusable autonomous-driving testbed based on San Francisco’s Market Street. The agent connected asset creation, traffic models, RTX sensor simulation and the Alpamayo driving stack. A subsequent Cosmos3-Nano experiment varied weather and lighting in recorded simulation videos to compare driving-model responses, according to the post.
Ashley Reid, who works on RTX sensor validation, directed Astra and Claude Fable 5 to compare simulated camera and lidar outputs with recorded data. The agents created two digital twins from scratch and improved two existing ones over about three days, iterating by measuring differences and adjusting OpenUSD scenes. The workflow addressed missing objects, geometry and materials, with acceptance guided by camera and lidar metrics, the blog post explains.
In the Robo Olympics project, Tae Kim, NVIDIA’s head of Omniverse engineering, used sports videos and natural-language instructions to have Astra build an experiment that tested simulated Unitree G1 humanoids performing sports movements. The agent constructed controllers and refined them with physics trials; the Newton Physics Engine simulated behavior, the Warp framework accelerated calculations and ovrtx rendered scenes. The robot cleared a single hurdle in 64 out of 100 simulation runs, the company reports.
The showcase also describes a robotic-disassembly workflow in which Astra modelled a car suspension in Onshape and designed a wrench for a robot to remove a suspension component in simulation, and an International Space Station application assembled from NASA assets with a single prompt. A further project turned stereo-camera captures into an editable OpenUSD studio for interaction testing.
Measured outcomes and missing verification
The Robo Olympics project reports a specific success rate (64 of 100 hurdle clears), and the digital-twin workflow describes a three-day iteration producing two new twins and improving two existing ones. Those details come from the blog post; NVIDIA provides no third-party benchmarks, replication data or comparisons with manual workflows. The company does not disclose iteration counts across the other projects, compute requirements beyond the three-day timeline, pricing or a timeframe for integrating the AI agents into products.
The post frames the showcase as a way for developers to see frontier AI models “at work” and promises further examples from NVIDIA teams and ecosystem partners.