Microduck open-source software: repositories, RL stack and limits

The two repositories do different jobs

The main microduck repository contains the onboard runtime and operational documentation: robot commands, gamepad pairing, updates, architecture and development setup.

The microduck_rl repository contains reinforcement-learning environments, simulation, PPO training and the export path used to run learned policies on the robot.

The train-to-deploy loop

Pollen’s published workflow trains movement policies in simulation, exports them to ONNX and runs them through the robot runtime. That split lets developers iterate in a virtual environment before risking physical hardware.

Commands, dependencies and branch guidance can change quickly during launch. Follow the current README and docs rather than copying a setup command from an undated guide.

Open-source software is not open-source hardware

The official press kit draws a clear boundary: the software stack is open, but the mechanical and electronic design files are not. Calling the whole robot open-source hardware would overstate the release.

• Open: runtime, SDK-related code, simulation and RL tooling.
• Not published as open hardware: mechanical and electronic design files.
• Check each repository’s current license and documentation before redistributing code.