Hermes, Nemotron, and Astra convert a prompt into a simulated obstacle course that Isaac Lab trains at scale.
NVIDIA published a workflow that uses AI agents to build and train a simulated obstacle course for Microduck, a reference robot. Hermes and Nemotron 3.5 Bolt on RTX Spark translate a prompt into a project plan, and Astra generates the 3D assets that populate the environment.
The pipeline spans RTX Spark as a local development hub, NVIDIA Isaac Sim for assembly, Omniverse libraries for rendering and sensor simulation, and Newton for physics modeling of movement and balance. Isaac Lab handles reinforcement learning across 4,096 parallel environments on an RTX PRO 6000 Blackwell Workstation Edition GPU.
Builders gain a single loop covering asset creation, physics, training, and local validation instead of gluing separate tools together. Isaac Lab trains policies at scale, and those policies return to RTX Spark, where teams check obstacle detection, movement adjustment, and course completion in Isaac Sim before touching hardware.
NVIDIA frames the workflow as preparation for sim-to-real transfer, pairing Better Actuator Models with domain randomization. Teams can pull ungated assets from the SimReady Central Library and follow the Microduck tutorial on Hugging Face to adapt the process to their own robots.
What matters
- NVIDIA agents Hermes, Nemotron, and Astra turn a single prompt into a full Microduck obstacle course.
- Builders can assemble assets, physics, and training in one pipeline instead of stitching tools.
- Isaac Lab trains 4,096 parallel environments on one RTX PRO 6000 before sim-to-real transfer.
Why it matters
Isaac Lab trains 4,096 parallel environments on one RTX PRO 6000 before sim-to-real transfer.
This GenAI News article was prepared in original wording using reporting and materials published by NVIDIA Developer Blog. Source reference: https://developer.nvidia.com/blog/using-ai-agents-to-teach-robots-like-microduck-to-walk-in-simulation/.
Drafted by the GenAI News review pipeline.