Microduck: The $399 Duck That Democratizes Physical AI

Hugging Face's most culturally significant launch this week isn't part of the $13B Nvidia deal — it's a $399 duck. Pre-orders for Microduck opened yesterday: a 25 cm open-source biped built with Pollen Robotics that waddles, picks up 800g with its beak, gets back up when it falls, and roller skates. Ships before Christmas.

The signal isn't the duck. It's the stack. Microduck carries a camera, lidar, and two IMUs. Every behavior is a policy trained in physics simulation and deployed straight to the real robot — sim2real you don't have to take on faith, you can run. The SDK, simulator, and full RL training stack are on GitHub, in the open. Retrain on your own machine, redeploy. That's the Raspberry Pi pattern applied to embodied AI: a consumer-priced on-ramp that turns "train a robot in simulation" into a weekend project instead of a research grant.

Two threads crossed this week. Anthropic's Model Hardware Standard gave agents a spec'd body for industry; Microduck gives everyone else a $399 one. Same loop — simulate, train policy, deploy — at opposite ends of the price curve. And while the headlines fixated on Nvidia buying the platform, HF CEO Clem Delangue spent the week selling open robotics against "black box" systems controlled by "a few organizations." The duck is the argument.

Verdict. If sim2real on a $399 biped holds up outside the lab, this becomes the cheapest embodied-AI research platform on the market — and a lot of people's first robot. Watch the GitHub repo, not the press release.