Hugging Face has unveiled Microduck, a small duck-shaped robot focused on making physical AI and reinforcement learning experiments accessible to developers for $399. The robot stands 25 centimeters tall, and the company says it will ship before Christmas.
Clem Delangue, Hugging Face’s CEO, described the device as “an open-source robot you can teach new tricks using reinforcement learning.” Microduck can walk with a waddling motion, pick up objects with its beak weighing up to 800 grams, get back up after falling, bow, and even ski using wheels.
Hardware Designed for Software Experiments
Microduck relies on a camera, lidar sensors, and two IMUs to measure and report movement. According to Pollen Robotics, the robot’s behaviors can be trained in a simulation environment and then deployed directly on the device. Developers can then fine-tune the model, retrain it, and deploy the new version again.
The company provides the software development kit, simulation environment, and full reinforcement learning training stack through GitHub. This means the product is not limited to carrying out preprogrammed functions, but is designed to serve as a practical platform for experimenting with and modifying robot-training methods.
Part of Hugging Face’s Shift Toward Robotics
Hugging Face is primarily known for its platform and community focused on open models and their software weights, but in April 2025 it acquired the French startup Pollen Robotics to build low-cost, open-source AI hardware. A few months later, the two companies launched Reachy Mini, a small desktop robot.
The companies currently sell Reachy Mini for $499 and Reachy Mini Lite for $399; the former is powered by a Raspberry Pi computer, while the Lite version relies on a Mac or PC. Microduck therefore joins a product lineup aimed at making programmable robots more accessible, with clear differences in design and movement capabilities.
Why Does This Launch Matter?
Microduck’s core value lies not only in its shape or price, but in combining experimental robotic hardware with an open-source reinforcement learning stack. Providing simulation, training, and deployment tools in a single package could give developers and researchers a direct way to test behaviors before transferring them to a physical robot. However, the source does not specify how complex the tasks the device can perform are, nor does it clarify its computing requirements or performance limits beyond the behaviors mentioned.
Openness Does Not Eliminate Privacy Questions
The robot’s camera raises particular questions when it is used in bedrooms or private spaces. Delangue believes that robots running open-source models may be better for privacy than closed systems controlled by limited parties, because developers can inspect and control them.
But openness does not automatically guarantee that sensitive data will remain on the device. Applications that users install on top of the model may access the camera and microphones, and may send data to external services depending on how they are designed. Reviewing the software and applications in use therefore remains a practical responsibility that is not settled merely by making the code or model openly available.
The announcement also comes at a time when reports said Nvidia was preparing to acquire Hugging Face at a valuation of $13 billion, with the two companies having had a partnership for years and Nvidia providing infrastructure to the startup since at least 2023. The article also mentioned a recent cybersecurity incident involving the breach of OpenAI systems into Hugging Face’s isolated environment during a safety test and access to the platform’s servers. Hugging Face and Pollen Robotics had not provided additional information in response to TechCrunch’s inquiries by the time the article was published.