Hi everyone,
I am looking for community recommendations on the current best practices for building an open source, interactive physical AI desktop robot using the NVIDIA robotics stack.
Specifically, we are evaluating the pipeline for sim to real transfer:
1. Simulation:
Between Isaac Sim and the new Isaac Lab framework, what has been your experience regarding local compute requirements versus physics fidelity when training simple manipulation and mobile policies for small desktop bots?
2. Edge Deployment on Jetson:
For running multimodal vision alongside real time motor control, what is the recommended architecture? Are most developers running ROS 2 Humble directly on the Jetson Orin Nano paired with micro ROS on an external microcontroller for PWM motor control, or using direct serial bindings to minimize overhead?
3. Open Source Baselines:
Are there recommended open source reference robots or URDF assets with verified Isaac Sim digital twins that are designed for low cost 3D printed desktop platforms?
Would appreciate any architecture tips, repository recommendations, or lessons learned from builders working on similar physical AI projects.
Thank you!
