Microduck is a 25 cm open-source bipedal robot designed for AI builders to train and deploy custom behaviors using reinforcement learning and simulation.
What the product does and how it is positioned
Microduck provides an open-source platform for robotics enthusiasts to experiment with reinforcement learning. The robot is designed to learn behaviors in a physics-based simulation environment before deploying those policies to the physical hardware.
The system includes a full software stack, including SDKs and training tools, allowing users to fork, retrain, and share their own behavioral policies with the community.
Source-supported ways to use the product
Users can develop and test new movement policies in a simulated environment before deploying them to the physical robot.
Developers can create and share unique robot behaviors, such as specialized gaits or interaction routines, with the broader community.
The documented workflow, where available
Develop and learn behavioral policies within a physics simulation environment on a local machine or via cloud services.
Transfer the trained policy from the simulation environment to the physical Microduck hardware.
Analyze performance, tune parameters, and re-train the model to improve behavioral outcomes.
Share the finalized behavioral policy with the community for collaborative development.
The Microduck ecosystem is built on a sim-to-real workflow, which bridges the gap between virtual training and physical execution. By utilizing the MuJoCo physics engine, users can iterate on complex movements without risking hardware damage.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
Yes, the SDK, simulation environment, and the full reinforcement learning training stack are open source and licensed under Apache-2.0.
Yes, every behavior is a policy that can be trained in a physics simulation on your own machine or via Hugging Face Jobs and then deployed to the robot.
The robot is equipped with a camera, LiDAR, and two IMUs to support its movement and interaction capabilities.
Yes, the robot ships with seven trained moves out of the box, including walking, sitting, standing, kicking, grabbing, and getting back up.
Users can monitor, configure, and update the robot using the provided robotctl command-line interface.
