Physical AI just got cheap enough to try

Building AI software is cheap — you pay for an API call and some compute. Building AI that actually works in the physical world has always needed expensive hardware most small teams could never justify. That barrier just dropped. Hugging Face, the open-source AI company, launched the Microduck: a 25-centimeter, duck-shaped robot you can pre-order for $399 and train yourself with reinforcement learning.

The Microduck is not a toy you watch from a distance. It is a real robotics platform — 15 motors, a camera, lidar (a sensor that measures distances with laser light), and two motion sensors. It walks, crouches, grabs objects with its beak, and rights itself when it tips over. It ships with seven pre-trained behaviors and can run new ones in real time. It even roller skates, if you buy the $39 accessory pack.

What makes it different is the software. The full training stack is open-source under the Apache 2.0 license, meaning anyone can read, change, or build on it. Developers write new behaviors inside a simulator — a free, software copy of the robot — and then push those behaviors to the physical device. Hugging Face calls it sim-to-real transfer. That is how professional robotics labs work. Now it runs on a $399 duck.

Hugging Face bought French robotics startup Pollen Robotics in April 2025 to build exactly this: hardware cheap enough that developers everywhere can experiment. The bet is that a community of builders will train the duck to do useful things — sorting mail, monitoring spaces, carrying small objects — and then share those trained behaviors openly. The real asset is the community building around it, not any single robot.

If your product or business might one day involve physical spaces — a warehouse, a clinic, a shop floor — now is a genuinely good time to understand how robot training works. You do not need a research lab. The starting price is $399, the code is on GitHub, and delivery is before Christmas.