microduck: A Rust Brain That Makes an 800g Duck Robot Walk with RL Policies

Control software for a small bipedal robot measuring 25cm and weighing 800g. It runs a 50Hz control loop on the Rockchip RK3566, driving 15 servos with reinforcement learning policies. It performs various actions such as walking, wheel driving, object grasping, and getting up after falling.

The project is structured as a single Rust workspace, with each function separated into an independent daemon, from robotd which handles the control loop to updates, Wi-Fi, Bluetooth, gamepad, and camera streaming. All clients call the same API via JSON-RPC contracts over Unix sockets.

Policy training is conducted in a separate repository, microduck_rl, using MuJoCo and PPO, and the exported ONNX files are loaded in this repository. Software updates are designed with signature verification and health checks to enable automatic rollback, reducing the risk of bricking the robot.