🦆 Put AI agents to work to help your robot learn to navigate. Using Microduck as the reference robot, this walkthrough shows how to create a simulated obstacle course with OpenAI GPT-6 Astra, build 3D assets with NVIDIA Omniverse libraries, simulate physics with Newton, train policies in NVIDIA Isaac... Lab, and validate them in Isaac Sim on NVIDIA RTX Spark.show more

@nvidiaomniverse Follow along and adapt the workflow to your own robot:

@OpenAI @nvidiaomniverse so glad Nvidia robotics finally moved beyond ROS way of working with robotics. even if ROS is used it's not a good path for modern iterative highly integrated policy based actuator control robotics. it's a middle layer that can fade away.

@OpenAI @nvidiaomniverse That’s really cool, but I want to 3D print my robotics parts, I’ll stick with SOARM-101.

@OpenAI @nvidiaomniverse Simulation to real pipeline is key for scalable robotics

@OpenAI @nvidiaomniverse Great walkthrough combining GPT-6 Astra, Omniverse assets, and Newton physics for robot navigation training is powerful.

@OpenAI @nvidiaomniverse Awesome, 12 layers of bleeding-edge AI stack and thousands of dollars in GPU compute so a simulated Microduck doesn't run headfirst into a virtual wall

@OpenAI @nvidiaomniverse Fun demo. Does the policy run zero-shot on the real Microduck, or did it need some fine-tuning on the hardware? Curious how much of the sim-to-real gap came down to Newton’s actuator models versus domain randomization.

@OpenAI @nvidiaomniverse Full loop from simulation to Isaac validation is super clean

@OpenAI @nvidiaomniverse SI’s real power isn’t just talking; it’s acting on the world. Using GPT-4o to steer Microduck in sim means we’re moving from static benchmarks to continuous learning loops. Practical progress over hype.

@OpenAI @nvidiaomniverse Incredible pipeline from simulation to real robot navigation done right

@OpenAI @nvidiaomniverse GPT-6 Astra helps create the obstacle course, while Isaac Lab trains the navigation policy. That separation makes Isaac Sim validation a distinct check on the learned behavior, rather than treating a generated course as proof the robot can navigate it.

@OpenAI @nvidiaomniverse Excellent pipeline simulation to real robot learning workflow

@OpenAI @nvidiaomniverse Nice walkthrough, thanks! Telling an agent to build the whole obstacle course and it just does it is exactly what I want to try on my own bench setup🤩

@OpenAI @nvidiaomniverse agents building the course, the assets and the training run .. great to see sim getting this easy !!

@OpenAI @nvidiaomniverse Incredible walkthrough using AI agents end to end from simulation to real navigation is next level!

@OpenAI @nvidiaomniverse Awesome end to end workflow for robotics learning

@OpenAI @nvidiaomniverse 从赛道生成到策略验证，整套链路串得挺顺