NVIDIA GPT-6 Astra Builds Seven Omniverse AI Simulations
NVIDIA shows GPT-6 Astra using Omniverse to build seven simulations spanning robots, autonomous driving, digital twins, space, captured rooms and CAD.
Summary
On October 8, 2026, NVIDIA detailed seven projects in which developers steer frontier agents with natural-language prompts, review outputs and request corrections, while Omniverse libraries supply GPU-accelerated physics, rendering and sensor simulation. Frank DeLise used GPT-6 Astra, SimReady, ovphysx, ovstage, ovrtx and ovui to turn a warehouse and humanoid into a gamified first-person and third-person simulator, with Astra generating animation and application code. Doyub Kim built Zero to Alpamayo on San Francisco's Market Street by connecting assets, traffic, Omniverse RTX sensors and Alpamayo driving. A Cosmos3-Nano experiment varied weather and lighting in recorded videos to compare responses. Ashley Reid guided Astra and Claude Fable 5 for about three days to compare ovrtx camera and raw LiDAR output with recordings, build two digital twins, improve two others, and revise OpenUSD objects, geometry and materials against camera and LiDAR metrics.
Tae Kim's experimental Robo Olympics used sports videos and prompts to make Astra build controllers for Unitree G1 humanoids. Newton Physics Engine, open-source NVIDIA Warp and ovrtx supported trials, with one robot clearing one hurdle in 64 of 100 runs. Jens Jebens had Astra model a car suspension in PTC Onshape, configure it in NVIDIA Isaac Sim, measure clearance and design a wrench that successfully removed a component in simulation, creating a starting point for robot policy training. Nic Johns used one prompt plus a daytime correction to combine NASA assets and telemetry into a browser-based OpenUSD International Space Station application using Blender, ovrtx, ovstage and ovstream. Chirag Majithia converted stereo captures into an editable OpenUSD studio with PyCuSFM, FoundationStereo, nvblox, generated and Blender assets, USD Content Agents, and Isaac Sim collision and contact tests for doors and drawers.
Positives
- Astra generated the animation and application code connecting four Omniverse libraries in Frank DeLise's warehouse humanoid simulator.
- Ashley Reid's agents created two digital twins and improved two others using measured camera and LiDAR discrepancies.
- A Unitree G1 humanoid cleared one hurdle in 64 of 100 Robo Olympics simulation trials.
- Astra designed a wrench that successfully removed a car suspension component in NVIDIA Isaac Sim.
- Nic Johns assembled NASA assets and telemetry into a browser-based International Space Station application with one initial prompt.
- Chirag Majithia turned stereo room captures into an editable OpenUSD environment for testing door and drawer interactions.
Risks & concerns
- The Unitree G1 cleared one hurdle in only 64 of 100 trials, leaving its timing and control open to improvement.
- Changes to scenes, sensors or driving models can significantly alter downstream autonomous-driving behavior.
- Digital-twin refinement required about three days of guided iteration and acceptance checks using camera and LiDAR metrics.
- Missing objects, geometry and materials created discrepancies between simulated sensors and recorded data.
- Developers still had to review agent output, direct revisions and validate each integration.
- The suspension component was removed only in simulation, providing a starting point rather than a deployed disassembly system.