MicroDuck

Microduck on GitHub: SDK, simulator and training code

Explore the official Microduck GitHub repositories: Python SDK, browser simulator, reinforcement-learning training code, licences and contribution paths.
Oct 22, 2026

Microduck robot beside a laptop and controller on a developer desk

Short answer: The official Microduck code lives at github.com/pollen-robotics/microduck, with reinforcement-learning environments in pollen-robotics/microduck_rl. The software, Python SDK, simulator and training stack are public under Apache-2.0. The mechanical and electronic design files are not published, and the simulation STL meshes use a separate CC BY-NC-SA 4.0 licence.

This page is the map from the storefront to the source. Start with the simulator or Python SDK if you want to use the robot, and use microduck_rl when you want to train a new policy.

Where is the official Microduck GitHub repository?

The official robot repository is:

https://github.com/pollen-robotics/microduck

It contains the robot software and documentation published by Pollen Robotics. The separate training repository is:

https://github.com/pollen-robotics/microduck_rl

If another repository copies the name but is not under the pollen-robotics organisation, treat its hardware and safety claims carefully. Forking open-source code is allowed, but a fork is not the official source of product support or current instructions.

What is open source in Microduck?

The published software stack is open source under Apache-2.0. That includes:

  • the robot runtime and command-line tools
  • the Python SDK
  • the simulation workflow
  • reinforcement-learning environments and training code
  • policy export and deployment tooling
  • technical documentation

The robot can walk without a subscription or cloud account. Policies run onboard, and the same software stack can be studied, modified and used as the basis for your own experiments.

The simulator is available as a hosted Hugging Face Space and can also be run locally from the published source. The Python SDK guide is the shortest path from a clone to a robot command.

What is not open source?

Microduck is not open-source hardware. The complete mechanical and electronic design files, manufacturing BOM and official assembly package are not published.

That distinction matters:

  • You can inspect and modify the software.
  • You can use the public simulation meshes for simulation and visual work.
  • You cannot treat the repository as a complete kit to manufacture another robot.
  • You cannot assume a printed part is mechanically validated just because a mesh is public.

The public simulation STL meshes are licensed under CC BY-NC-SA 4.0, separately from the Apache-2.0 software. Read the repository's licence files before redistributing a model or using it commercially. The 3D print guide explains the actual limits of the published meshes.

Which Microduck repository should I start with?

Choose by the job you want to do.

GoalStart hereWhy
Meet the robot without hardwareBrowser simulatorNo install and no robot required
Send a first Python commandPython SDK guideConnects the published SDK to a practical first session
Understand the runtimeSoftware architectureExplains daemons, IPC and the 50 Hz control loop
Train a policyTraining guideWalks through microduck_rl, mjlab and sim-to-real
Install or publish a policyPolicies guideCovers slots, skills, ONNX and manifests
Build or print for the robotDIY guideSeparates published simulation assets from hardware designs

The robot repository is the source of record for runtime behavior. The training repository is the source of record for training environments. Avoid copying commands from an old blog post when the same command is documented in the current repository.

How do I run the simulator or Python SDK?

The fastest simulator path is the hosted Hugging Face Space linked from the simulator page. It lets you drive the robot and inspect behaviours before hardware arrives.

For local development, clone the repositories in the order the task requires:

git clone https://github.com/pollen-robotics/microduck
git clone https://github.com/pollen-robotics/microduck_rl

Read the current setup instructions in each repository before installing dependencies. Hardware, CUDA and Python requirements change more often than the overview pages on a storefront. The Documentation index is the maintained entry point for the robot runtime, simulator and SDK.

The standard robot kit includes the controller, battery and USB-C cable. A computer is useful for Python and simulation, but it is not required to drive the robot with the included controller.

How do I train and publish a Microduck policy?

Training happens in simulation, not on the robot. The published loop is:

  1. Run the environment in microduck_rl.
  2. Train a policy with the documented reinforcement-learning workflow.
  3. Export it through the official exporter.
  4. Test in simulation and with the robot supported safely.
  5. Load it into an appropriate slot or publish it as a skill.

The full workflow is in the training guide, and the install/export details are in the policies guide.

Local training benefits from a CUDA GPU. Hugging Face Jobs is the documented hosted-compute alternative. The robot can run a finished policy without a GPU; the GPU is for training, not for everyday control.

How can I contribute or report a bug?

For a reproducible software issue, use the repository's issue tracker and include:

  • robot or software version
  • operating environment and installation method
  • exact command that triggered the issue
  • expected result and actual result
  • relevant logs with secrets removed

For a pull request, keep the change focused and follow the contribution guidance in the repository. Hardware safety boundaries matter: only the motor-owning daemon should command the hardware, and a change that bypasses those boundaries should not be treated as a routine contribution.

For a classroom project, a policy experiment or a general question that is not a reproducible bug, the Pollen Robotics Discord is the community route.

Is Microduck GitHub code safe to run?

Open source means the code can be inspected; it does not mean every fork or community policy has been reviewed. Use the official repository for current setup instructions, read the code before connecting hardware, and test new policies with the robot on a stand and away from table edges.

Community policies may have their own provenance and licence. The policies guide explains how the robot checks policy shape and why a successful simulator run is not a substitute for safe hardware testing.

Microduck GitHub FAQ

Where is the official Microduck GitHub? The official repository is github.com/pollen-robotics/microduck. Training code is in github.com/pollen-robotics/microduck_rl.

Is Microduck fully open source? The software stack is open source under Apache-2.0. Complete mechanical and electronic design files are not published, so it is not open-source hardware.

Can I clone the repository and build a robot? No. You can inspect and modify the published software, but the repository is not a complete manufacturing package.

Do I need a GPU to use the SDK? No. A GPU is useful for training policies, not for driving the robot or sending ordinary SDK commands.

Where should I ask for help? Use the troubleshooting guide and repository issues for reproducible bugs, Discord for community help, and support@microduck.homes for order or hardware-specific questions.

Key takeaways

  • pollen-robotics/microduck is the official robot software repository.
  • pollen-robotics/microduck_rl is the reinforcement-learning repository.
  • Apache-2.0 covers the software; it does not mean the hardware design is public.
  • Simulation STL meshes have a separate CC BY-NC-SA 4.0 licence.
  • Start with the simulator and Python SDK before moving into policy training.

Where to go next

Microduck GitHub: SDK, simulator and RL code | MicroDuck