Hugging Face’s $399 rollerskating duck robot

What happens when the biggest hub for open-source AI decides to build a robot? You get a tiny duck that roller-skates across your living room floor.

Hugging Face and its robotics arm, Pollen Robotics, just unveiled Microduck — an adorable, open-source bipedal robot that stands under 10 inches tall (about 25 cm) and weighs just 800 grams. Priced at $399, it is available for preorder in cream, graphite, lavender, and sky blue, with shipments scheduled to arrive before Christmas 2026.

A tiny biped with a one-of-a-kind voice

Microduck may look like an oversized desk toy, but its mobility suite is remarkably versatile. Official demos show the one-eyed duck waddling around, chasing laser pointers, and bending over to pick up socks and markers with its beak. Put on a pair of tiny wheels, activate roller mode, and it swaps policies to skate around. If you knock it over, it stands itself back up.

Instead of acting like a polite virtual assistant, Microduck acts more like a pet creature. The very first time it boots up, the robot generates an audio identity of weird chirps and quacks that stays with that specific unit for life. Users can drive it around with a regular game controller or manage it locally from a laptop over Bluetooth without touching SSH.

Open-source reinforcement learning under the hood

Beneath its colorful shell, Microduck is a fully fledged physical AI testbed. It runs on a Rockchip RK3566 processor driving 15 servos at a 50 Hz control loop written in pure Rust. It perceives the room via a camera, LiDAR, and two inertial measurement units (IMUs).

Crucially, all the motion control runs on neural policies trained with PPO inside the MuJoCo physics simulator and exported to ONNX. Hugging Face released the entire training stack, SDK, and simulation code on GitHub, meaning developers without the physical hardware can still spin up simulated ducks on their laptops to train new tricks and deploy them later.

The release follows Pollen's earlier Reachy Mini robots ($399 to $499) and arrives amid reports that Nvidia is looking to acquire Hugging Face in a $13 billion deal.

Why it matters

We think Hugging Face nailed the wedge for physical AI. Training reinforcement learning on actual physical bipeds usually requires tens of thousands of dollars and specialized lab environments where a single broken actuator ruins your week. At $399, Microduck delivers a turnkey, end-to-end Sim2Real pipeline that any student or indie hacker can break and tinker with on their desk.

It also marks a shift away from over-hyped, full-sized humanoids. Instead of promising an autonomous butler that folds laundry, Hugging Face is doing what it does best: open-sourcing the primitives and letting the developer community crowdsource locomotion and interaction policies.

Still, buyers should keep expectations in check. The RK3566 is an entry-level chip, meaning complex onboard vision-language models won't run locally without external compute. There is also an inherent privacy tension: while open-source firmware prevents corporate backdoors, loading third-party community apps onto a roaming bot equipped with cameras and microphones in private spaces could easily become a security headache.