Train in Simulation, Deploy on a Real Robot: Microduck with Isaac Lab 3.0
huggingface ·
This article introduces Microduck, an affordable open-source bipedal robot developed by Pollen Robotics and Hugging Face, demonstrating how to train it for locomotion using NVIDIA Isaac Lab 3.0. Developers can utilize Isaac Lab's GPU-parallel simulations and Newton-based actuator modeling to create robust walking policies that account for real-world factors like friction and mechanical backlash. The tutorial guides users through installing the framework, configuring environments for flat or rough terrain, and accelerating training across multiple GPUs. Once trained, policies are evaluated in simulation and exported as ONNX files for deployment on the physical hardware. Users can share their models on the Hugging Face Hub or test them in the Microduck Arena competition space. The guide details the specific steps for transferring the ONNX policy to the robot via SSH and loading it into the onboard runtime. Finally, the article suggests extending these capabilities to more complex tasks such as perception-based navigation and dexterous manipulation.
huggingface