Hugging Face Unveils Microduck: A $399 Open-Source 25 cm Biped You Train with Reinforcement Learning

What it does

Pollen Robotics, working with Hugging Face’s Bordeaux team, launched Microduck, a 25 cm tall bipedal robot that costs $399. It features 15 motors, an onboard camera, LiDAR, two inertial measurement units, and a full software stack for reinforcement learning. Each movement of Microduck is driven by a neural policy trained using the MuJoCo physics simulator and exported to the ONNX format. Buyers get a complete sim-to-real loop on their desk, including open-source training tools under an Apache-2.0 license, allowing users to retrain and customize the robot’s behavior.

Why it matters

Microduck packages advanced robotics and AI training into an accessible, affordable device for developers and researchers. It lowers the barriers to experimenting with real-world reinforcement learning by integrating sensors and actuators with an open-source training stack. Unlike typical robots requiring expensive hardware and cloud services, Microduck puts the entire development cycle locally, speeding iteration and reducing dependency on proprietary systems. This approach shifts control and cost down to the developer level, making robotics research and product development more hands-on and affordable.

Who it is for

This robot primarily targets robotics developers, AI researchers, and educational institutions looking to experiment with bipedal locomotion, sensor fusion, and reinforcement learning. It also appeals to startups and innovators testing robotics applications without investing heavily in hardware or cloud infrastructure. The open training stack enables advanced users to tweak and retrain policies, making it a testbed for custom movement algorithms and real-time sensor data integration.

The catch

While $399 is low for a bipedal robot, Microduck’s size and payload limits mean it remains a prototype platform rather than a production-ready system. The reliance on MuJoCo—a commercial physics engine—adds some dependency outside the open-source stack, although training export to ONNX maintains flexibility. Developers need familiarity with reinforcement learning workflows and robotics sensors to extract full value, which could narrow accessibility for beginners.

What to watch next

Microduck could pressure other robotics makers to provide more integrated, affordable training and simulation environments. Tracking how the community adopts or extends the open-source training stack will reveal whether this local sim-to-real model scales beyond research. Watch for integrations with other AI frameworks, expanded sensor suites, or larger form factors that bring similar packages to industrial or consumer markets.

AI Quick Briefs Editorial Desk