Pollen Robotics and Hugging Face release Microduck, a walking biped trained in simulation

From simulator to physical robot

The Microduck workflow has three stages: a walking policy is trained in simulation, exported in ONNX format and deployed on a real biped robot. According to the developers, a policy trained in simulation must also cope with the friction, delays and mechanical play found in real robot servos, which is why the approach uses actuator modelling and domain randomisation.

NVIDIA Isaac Lab, an open-source framework for training policies across thousands of GPU-parallel simulations, provides the training environment. Isaac Lab already supplies the Microduck robot model together with its scene, observations, actions, rewards and domain randomisation settings. Users begin with a flat-walking task, while a rough variant adds stairs, slopes and uneven terrain.

Requirements for training and deployment

Simulation and training call for a Linux workstation with a supported NVIDIA GPU and driver, plus an Isaac Lab installation. Deployment also requires a configured physical Microduck with its onboard robot software installed and working, and a supported control interface such as a paired gamepad.

Isaac Lab runs simulated environments in parallel on the GPU to speed up locomotion training, and it can distribute training across GPUs, for instance across 16 GPUs.

Next steps for developers

A trained and evaluated policy can be shared on the Hugging Face Hub, checked for compatibility and entered in the Microduck Arena competition. Developers can also run it on their own Microduck or move on to more advanced tasks with Isaac Lab 3.0, for which linked examples are offered.

Open questions

• How well do trained walking policies perform on physical Microduck robots outside the tutorial setup?