Hugging face's microduck: a compact open-source bipedal robot for reinforcement learning

Hugging Face's Pollen Robotics has open-sourced Microduck, a 25cm-tall, 800g bipedal robot designed for reinforcement learning and real-time control. The robot integrates advanced technologies such as Sim2Real transfer and ONNX deployment, enabling developers to train and deploy models on real hardware. Microduck features a Rockchip RK3566 processor and 15 servos, allowing it to walk, stand, pick up objects, and perform actions like kicking a ball and doing a forward roll. It can also switch to a wheeled mode for sliding and turning. The training project, microduck_rl, uses MuJoCo and PPO algorithms with domain randomization to bridge simulation and reality. The software, written in Rust, includes safety checks and a robust update process to prevent system failures. The robot's open-source nature provides a complete pipeline for research in embodied intelligence and robot motion control.