On August 27, Hugging Face announced the Microduck: a 25-centimetre, $399 open-source robot duck that ships before Christmas 2026. It is genuinely cute, which is doing a lot of marketing work — but look past the duck and this is one of the most accessible serious robotics platforms anyone has put on sale.
What Is Actually Inside
For $399 you get a sensor package that would have been a research budget line item a decade ago:
- ›A camera for visual perception
- ›Lidar sensing
- ›Two IMUs (inertial measurement units) for balance and orientation
- ›Onboard compute driving it all
The demonstrated behaviours are a good indication of what that sensor stack supports: it waddles, picks objects up with its beak, crouches, recovers from falls, and roller skates. Fall recovery in particular is not a gimmick — it is a legitimately hard closed-loop control problem, and shipping it as a stock behaviour says something about the quality of the underlying stack.
The Real Product Is the Training Stack
The hardware is the hook. The reason this matters is that Hugging Face put the SDK, the simulation environment, and the full reinforcement learning training stack on GitHub. Behaviours are trainable in simulation and deployable directly to the physical robot.
That sim-to-real loop is the part that has historically gated people out of robotics. Building a physics simulation, wiring it to a training pipeline, and then closing the gap to real hardware is months of infrastructure work before you learn anything. Handing someone a working loop and a $399 robot at the end of it collapses that barrier almost entirely.
The designs and the code are both open, which means you can audit what the robot does, change it, and retrain it — the difference between owning a robot and renting one.
How It Fits the Wider Hugging Face Robotics Push
This did not come out of nowhere. Hugging Face acquired the French robotics startup Pollen Robotics in April 2025 and has been shipping hardware since, including the Reachy Mini desktop robot at $499 (Raspberry Pi-powered) and the Reachy Mini Lite at $399 (driven by a connected Mac or PC). CEO Clem Delangue has framed the whole effort in terms of democratising physical AI and world models — taking the same open-weights posture the company took with language models and pointing it at robots.
Why This Is Interesting for Final-Year Projects
We see a lot of robotics capstones that stall in the same place: the team builds a chassis, gets it to drive in a straight line, and runs out of semester before doing anything that demonstrates learning. A platform where locomotion already works and the training pipeline is handed to you moves the starting line forward by months. The interesting work — reward design, sim-to-real transfer, teaching a new behaviour — becomes the whole project instead of the part you never reach.
The Caveats Worth Naming
Hugging Face is explicit that the Microduck is aimed at developers and researchers interested in physical AI and world models, not at consumers — expect rough edges, not a polished toy. It also does not ship until late 2026, so nobody outside the company has independently tested those behaviours yet. Judge it on delivery. But as a statement about who gets to do robotics research, $399 is a meaningfully different number than what came before it.
Written by Manas Garge
Founder & Data Engineer