Microduck: Tiny 25cm Duck Robot Uses Reinforcement Learning to Walk, Grasp, and Quack

Meet Microduck, a palm-sized 25cm biped duck robot that walks, grasps objects, and quacks using reinforcement learning — and its open-source design has already earned over 1,000 GitHub stars.

• Microduck is a tiny 25 cm, 800 g biped duck robot that uses reinforcement learning policies to walk, roll, grasp objects, stand itself up, and even quack in its own unique voice.
• The robot runs on a Rockchip RK3566 processor, powered by a suite of Rust-based daemons handling a 50 Hz control loop, Bluetooth, camera streaming, depth sensing, and over-the-air updates with automatic rollback if the robot comes up unhealthy.
• The open-source repository includes everything needed to run a Microduck, with policies trained using MuJoCo and PPO in a companion repo, exported to ONNX, and the project has already attracted over 1,000 GitHub stars and 98 forks.