Name: Microduck
Brand: Hugging Face

The Microduck does not look like a lab instrument. It is 25 centimetres tall, shaped like a cartoon duck, and ships with wheeled feet so it can roller skate. What separates it from anything you could buy before at this price is that it accepts new behaviors you trained yourself — on an ordinary home computer, using the same class of machine learning that Google DeepMind uses to teach robotic hands to stack objects in research facilities.

The robot’s open-source software stack covers simulation, robot control, and full reinforcement learning training. A developer writes a behavior in simulation, runs the training on their laptop, and transfers the result directly to the physical hardware. Hugging Face calls it sim-to-real deployment; it used to require specialized infrastructure and equipment costing tens of thousands of dollars.

The physical capabilities are modest. Fifteen motors drive a waddling gait and an articulated beak capable of picking up lightweight objects — a sock, a marker — and placing them in containers. A front-facing camera, a LiDAR sensor, and two inertial measurement units give the robot a picture of its surroundings. It can self-right when knocked over. The default behaviors feel less like a research demo than a well-designed toy — the $399 price matters because the machine’s value is what you can teach it, not what it already knows.

Comparable platforms sold to universities and corporate R&D teams run from $20,000 to $150,000 and typically come with proprietary software that limits modification. The Microduck’s full training pipeline lives on GitHub under an open-source license, which means the community can extend it without waiting for Hugging Face to ship updates. The company built the platform after acquiring Pollen Robotics, a Bordeaux-based robotics studio, in April 2025.

The open-source framing needs a caveat. The software stack is fully open; the hardware is not. Pollen Robotics has confirmed it has no plans to release hardware blueprints. The beak’s lifting capacity tops out at roughly 100 grams — the weight of a large apple — and features the company describes on the roadmap, including autonomous navigation and more dexterous grasping, depend on community contributions that do not yet exist. Training new behaviors also requires comfort with Python and a basic understanding of reinforcement learning. This is not a consumer device. It is an entry-level research tool for developers ready to move from software models to physical deployment.

Preorders are open internationally at $399, with Hugging Face targeting delivery before December 25, 2026. Regional distribution partnerships and simulation system requirements have not yet been detailed. Whether the price unlocks what it promises depends on a developer community that has not yet started building — and on hardware blueprints that Hugging Face controls and has not released.

Tags: Hugging Face, Open Source, robotics, Microduck, Pollen Robotics, reinforcement learning