Microduck can be trained with Isaac Lab 3.0 and deployed via ONNX

The guide walks users through installing Isaac Lab on a Linux workstation, launching the rough‑terrain training task with up to 1.25 M steps per second on 16 GPUs, and exporting the resulting policy as both TorchScript and ONNX files. The ONNX model (max 64 MiB) can be loaded onto the robot’s onboard runtime, tested in the Hugging Face‑hosted Microduck Arena, or shared on the Hub for community reuse. The article also lists the hardware requirements, command‑line steps, and how to handle servo backlash and other real‑world effects during simulation.

By combining open‑source hardware, a high‑performance simulation stack, and a straightforward deployment pipeline, developers can iterate quickly from virtual training to a physical robot that recovers from falls and walks over uneven terrain.