Agents

NVIDIA Automates 3D Scene Prep With GPT-6 Astra

NVIDIA has detailed a new agentic AI workflow using GPT-6 Astra and Omniverse libraries to automate the tedious process of preparing 3D Blender scenes for robotics simulation.

NVIDIA Developer Blog14 hrs agoAgents
Image: NVIDIA Developer Blog

NVIDIA has unveiled an agentic AI workflow designed to automate the labor-intensive process of preparing 3D art assets for physical AI simulations. Coordinated by an orchestrator like Codex powered by OpenAI's GPT-6 Astra or Anthropic's Claude, the system automates tasks that typically bottleneck robotics training. The workflow bridges the gap between creative 3D environments like Blender and simulation platforms like Isaac Sim or Isaac Lab, ensuring that virtual worlds are physically and semantically ready for virtual robot training.

The architecture relies on NVIDIA NemoClaw to deploy specialized Hermes subagents, which are powered by NVIDIA Nemotron models. These agents interact with Blender through a Model Context Protocol server to inventory the scene. They then write data directly to OpenUSD, a format that preserves the original scene hierarchy and metadata. To make the environment simulation-ready, the subagents use specialized Omniverse tools: ovphysx for applying collision meshes and rigid-body physics, ovrtx for rendering preflight views to check sensor visibility, and SimReady validation to ensure the scene meets strict simulation standards.

For robotics engineers, this automation eliminates the tedious manual labor of tagging semantic labels, configuring lidar and camera sensors, and debugging physics errors. Instead of manually adjusting every virtual object, developers can let the AI agents automatically repair minor issues, while complex or ambiguous cases are flagged for human review. This drastically accelerates the pipeline for creating digital twins, allowing engineers to focus on training policies rather than fixing broken 3D files.

To run these agentic pipelines, NVIDIA recommends several hardware tiers. Developers can prototype locally on the NVIDIA DGX Spark, which features 128GB of coherent unified system memory. For more demanding local workloads, the NVIDIA DGX Station offers up to 748 GB of coherent memory via the GB300 Grace Blackwell Ultra Desktop Superchip, with the option to add an RTX PRO 6000 Blackwell GPU. Larger teams can scale up using RTX PRO Servers or DGX Cloud to handle massive batch scene validation and synthetic data generation.

This is our own summary of reporting by NVIDIA Developer Blog

More in Agents