Microduck

Microduck is a 25 cm biped robot designed to make physical AI and reinforcement learning approachable on a desk or workbench. It has 15 motors, a front camera, compact LiDAR, two IMUs, microphones, a speaker, and an articulated beak for grasping objects. The robot weighs under 800 grams and includes a game controller, so it can be played with before any code is written. Out of the box, Microduck can walk, sit and stand, kick, grab, roller-skate when equipped, and recover from several common falls. Developers can then train new behaviors in a physics simulator on their own machine or through Hugging Face Jobs, deploy the learned policy to the physical robot, refine the simulation, and publish the result for others to use. The workflow is built around sim-to-real transfer rather than repeated trial-and-error directly on hardware. The open-source software stack includes the SDK, virtual training environments, reinforcement-learning scripts, robot-control tools, and deployment utilities. Microduck is intended for robotics developers, educators, experimenters, and learners who want a small, playful platform for exploring movement, recovery, grasping, and real-world AI behavior.