(VIDEO) Hugging Face Unveils Microduck, a $399 Open-Source Robot Duck You Can Teach New Tricks Right at Home

SAN FRANCISCO — Hugging Face, the AI platform known primarily as a hub for open-source software models, unveiled a small robotic duck Thursday that developers, students and hobbyists can train at home using reinforcement learning, marking the company’s latest push into affordable, open-source physical AI hardware.

The device, called Microduck, is now available for pre-order at $399, with shipments expected before Christmas, according to TechCrunch. The launch was announced by Hugging Face co-founder and CEO Clément Delangue on social media Thursday morning.

A tiny robot built for experimentation

Standing just 25 centimeters, or roughly 9.8 inches, tall and weighing about 800 grams, or under 2 pounds, Microduck is designed less as a finished consumer gadget and more as a hands-on platform for people interested in experimenting with reinforcement learning and what the industry calls “physical AI” — systems that learn to interact with the real world rather than operating purely in software.

Delangue described the device in direct terms on social media. “It’s a tiny $399 open-source robot you can teach new tricks with reinforcement learning,” Delangue said, according to TechCrunch.

What it can do out of the box

Microduck ships with seven pre-trained behaviors, according to a report from Tao Media: it can walk, sit, stand, kick, grab small objects, roller-skate and recover to its feet after falling over. The robot’s articulated beak doubles as a small gripper, allowing it to pick up objects weighing up to 800 grams. A game controller is included with the device to allow immediate hands-on interaction.

Beneath its playful design, Microduck is equipped with a substantial sensor package for a device at this price point. According to Tao Media, the robot packs 15 motors along with a wide-angle camera, LiDAR, microphones, a speaker, two inertial measurement units, near-field communication, Wi-Fi and Bluetooth connectivity.

Open source from top to bottom

Central to Microduck’s design is its fully open software stack. The robot’s software development kit, its MuJoCo-based simulation environment, and its full reinforcement learning training stack are published publicly on GitHub under an Apache 2.0 license, according to Tao Media. Even the seven behaviors that ship with the device can be inspected and retrained by users, meaning nothing about how the robot operates is hidden inside a proprietary black box.

That openness extends to how users can develop new skills for the robot. Because training reinforcement learning models directly on physical hardware would require thousands or even millions of trial-and-error attempts, likely damaging the robot in the process, Hugging Face and its robotics partner have built the system around a “sim-to-real” workflow, according to Hackster.io. Developers train new behaviors virtually in simulation first, then transfer the resulting policy onto the physical robot. Users can run that training locally on their own computers or through Hugging Face’s cloud infrastructure before deploying it to the device.

Part of a broader robotics push

Microduck is not Hugging Face’s first venture into physical hardware. The company acquired French robotics startup Pollen Robotics in April 2025 specifically to build affordable, open-source AI hardware, and the two companies previously launched Reachy Mini, a small desktop robot, months after the acquisition, according to TechCrunch. Hugging Face had also previously released open-source humanoid robots called HopeJR and Reachy Mini at a lower $250 price point.

The broader strategy reflects a belief within the company that robotics represents the next major frontier for artificial intelligence, extending a technology that has largely been confined to software into systems capable of physically interacting with the world, according to Axios.

Addressing privacy concerns

The launch of a camera-and-microphone-equipped robot designed for home use raises familiar privacy questions that have followed other AI hardware devices. Delangue has previously addressed those concerns directly with TechCrunch, arguing that robots powered by open-source models offer better privacy protections than what he described as “a black box system” controlled by a small number of organizations — a distinction he said matters especially when the leadership of those organizations lacks stability.

Even so, open-source design does not guarantee complete data privacy on its own. Once consumers install third-party software applications on top of an open model, those apps can gain access to a device’s cameras and microphones, and depending on how they are built, could potentially transmit that data to outside services.

Launch comes amid acquisition speculation

Microduck’s debut arrives at a notable moment for Hugging Face. The company is reportedly set to be acquired by Nvidia at a valuation of roughly $13 billion, according to a report from The Information cited by TechCrunch. Nvidia and Hugging Face have maintained a partnership for years, with Nvidia providing infrastructure support for the startup’s operations.

When asked about the reported deal, Delangue declined to comment directly on reporting that the company had hired bankers to explore a potential sale valued near $13 billion, according to Axios, and did not immediately respond to questions about the specific report of a possible acquisition by Nvidia.

Lowering the barrier to robotics research

The broader significance of Microduck, according to those covering its release, lies in its price point relative to the cost of comparable research robots. Much of the current wave of interest in physical AI has centered on expensive humanoid robots and industrial systems that remain largely inaccessible to individual developers, students and hobbyists. By packaging core concepts like embodied perception, reinforcement learning and sim-to-real policy deployment into a $399 device, Hugging Face and Pollen Robotics are betting they can meaningfully broaden who gets to participate in robotics experimentation, much as the company’s model-sharing platform has done for software-based AI development over the past several years.