Microduck is a tiny biped robot that moves using reinforcement learning policies.
The robot solves the problem of creating a small, agile platform capable of complex locomotion and manipulation tasks. It runs on a Rockchip RK3566 processor and executes neural network policies trained with reinforcement learning, maintaining a 50 Hz control loop that drives fifteen servos. The approach combines sim-to-real transfer learning with ONNX model export, allowing policies trained in simulation to run directly on the hardware. The robot can walk under gamepad control, roll on wheels by switching policies, pick up objects with its beak, recover from being knocked over, and perform additional behaviors like sitting, kicking, and vocalizing.
Adoption suits developers and roboticists interested in embodied AI, reinforcement learning applications, or small-scale robotics platforms. The tool works well for projects requiring dynamic locomotion or manipulation with learned behaviors rather than hand-coded control. The repository provides the complete runtime system needed to operate a Microduck, including daemon processes for servo control, camera and radio handling, and robust update machinery that prevents bricking during software installation. Separate documentation covers gamepad pairing, command-line control via duckctl over Bluetooth, and reversible update procedures with health gating.
The project maintains active continuous integration and provides clear architectural documentation. Development activity shows ongoing maintenance of the core control systems and update infrastructure, with established patterns for configuration and troubleshooting documented in the cheat sheet and specialized guides.
