Hugging Face and NVIDIA collaborators published a tutorial for Microduck, an affordable open-source bipedal robot from Pollen Robotics and Hugging Face. The workflow trains a walking policy in simulation with Isaac Lab 3.0 and the Newton physics engine, then exports it as ONNX for deployment on the physical robot. Walking is presented as a starting point, with linked examples for exploring more advanced Isaac Lab tasks.
The tutorial provides a ready-to-train Microduck environment with both flat and rough terrain variants; the rough task adds stairs, slopes, and uneven terrain. The policy observes 61 values and outputs 14 joint positions. To help the policy transfer from simulation to hardware, domain randomization varies mass and center of mass, ground friction, joint inertia, encoder offsets, battery voltage and sag, command latency, terrain, pushes, and gearbox backlash. Backlash is modeled per joint because the mechanical play in a gearbox can differ across joints as parts wear unevenly.
Training runs many simulated environments in parallel on NVIDIA GPUs, with the example using 4,096 environments and the RSL-RL library. With 16 GPUs, the authors report reaching up to 1.25 million steps per second on rough terrain with backlash enabled. The playback workflow exports both TorchScript and ONNX files, and the ONNX policy is intended for Microduck's onboard runtime. Users can also validate policies in the Microduck Arena or share them on the Hugging Face Hub.