NVIDIA documents the path from simulated walking to a real Microduck
Its October 21 tutorial connects Isaac Lab training, simulation checks and an exported walking controller for the physical robot.
NVIDIA’s tutorial describes training a walking controller in simulation, evaluating it, then exporting an ONNX model file for Microduck. Hardware deployment needs a configured robot with working onboard software and a supported controller, such as a paired gamepad. The observation order, normalization and action scaling must stay consistent with training: exporting a file alone does not complete the documented handoff.
The installation guide clarifies the computer requirements behind that workflow. Isaac Lab installs the Newton physics system by default; Isaac Sim is optional. The default training environment needs a supported NVIDIA GPU, with the CUDA 13 stack requiring Turing or newer hardware. Memory needs depend on the task, parallel simulations and camera workload. Newton simulation without a display does not require RTX hardware, but adding a renderer or interactive viewer adds its own requirements. The recommended uv setup manages the Python environment.
The robot’s runtime guide explains what replacing the walking controller actually changes. If Microduck is walking, loading a replacement returns it to its home pose, loads the model and resumes motion. If it is sitting or standing still, the replacement does not move it. The saved choice survives reboots and software updates. A model with the wrong input or output shape is rejected before changes; a failed load retains the previous controller. Resetting the walking slot restores its default policy.