NVIDIA uses AI agents to build Omniverse simulations

Developers are pairing frontier AI models including GPT-6 Astra with NVIDIA Omniverse libraries to move simulation ideas into working applications. The workflows use natural-language prompts while Omniverse supplies GPU-accelerated physics, rendering and sensor simulation, with developers reviewing outputs and steering changes.

NVIDIA projects show Astra connecting Omniverse components for a SimReady warehouse humanoid simulator, building Zero to Alpamayo for autonomous-driving tests on San Francisco’s Market Street, and assembling an OpenUSD International Space Station browser app with NASA assets and telemetry. Other work used recorded camera and raw LiDAR outputs to improve digital twins, with agents creating two scenes and refining two existing ones over about three days.

Robot-focused experiments included Robo Olympics, where simulated Unitree G1 humanoids learned sports movements through controllers tested with the Newton Physics Engine, NVIDIA Warp and ovrtx; one hurdle experiment succeeded in 64 of 100 simulation trials. Additional work modeled robotic disassembly of a car suspension in PTC Onshape and Isaac Sim, while captured-room workflows converted stereo camera data into editable OpenUSD spaces for interaction testing.