Hugging Face's robotics arm Pollen Robotics has launched the Microduck — a $399, one-eyed biped robot standing just under 10 inches tall that owners can train with reinforcement learning — and it ships before Christmas 2026.
Key takeaways
The Microduck is priced at $399 and available to preorder now from Pollen Robotics, with shipping planned before Christmas 2026.
It stands just under 10 inches tall and comes in four colors: cream, graphite, lavender, and sky blue.
The robot is fully open-source; Hugging Face CEO Clem Delangue describes it as a robot "you can teach new tricks with reinforcement learning."
It is Pollen Robotics' second consumer robot, following an earlier, larger model.
The Microduck's open-source design means hardware schematics and software are intended to be publicly accessible and modifiable.
Three Microduck robots in different colorways play-kicking a ball — a demo of the robot's locomotion capabilities.
The Microduck is a compact thing: one cyclopean eye, two stubby legs, and a silhouette that lands somewhere between a bath toy and a Bauhaus sculpture. Pollen Robotics — the French robotics company Hugging Face acquired in 2024 — built it as a platform for hands-on machine learning experimentation, not just a novelty. The reinforcement learning hook is the real pitch: owners can define reward signals and let the robot figure out behaviors on its own hardware, closing the loop between simulation and physical motion in a way that most consumer robots never attempt.
What the open-source architecture actually enables
For AI practitioners who have spent time training models in software, the Microduck offers a rare chance to transfer that intuition into physical space. Reinforcement learning on a $399 body is a different proposition than spinning up a cloud GPU — the feedback loop is slower, the failure modes are visible and sometimes funny, and the constraints of real-world physics impose a discipline that purely digital environments don't. That friction is also a creative constraint: robot locomotion trained from scratch tends to produce movement that is genuinely strange and hard to predict, which has its own aesthetic interest.
"Open-source robot you can teach new tricks with reinforcement learning."
— Clem Delangue, CEO of Hugging Face
The open-source framing matters practically. Hugging Face has built its reputation on accessible model weights and community-driven iteration — the same logic that powers the Hugging Face Gradio visual workflow builder for multi-model AI pipelines. Applying that philosophy to hardware means the Microduck's schematics and codebase are meant to be forked, modified, and improved by the community, not locked behind a proprietary SDK. Whether Pollen Robotics delivers on that openness at launch is something preorder buyers will be watching closely.
The Microduck in lavender colorway, showing its single-eye design and compact biped form factor.
According to The Verge, the Microduck is available in cream, graphite, lavender, and sky blue — color choices that read less like industrial robotics and more like a consumer electronics launch calibrated for a specific kind of enthusiast. The rollerskating demos circulating online lean into the same register: this is a robot designed to be charming first, capable second, and experimentable third.
For AI creators whose work lives entirely in image generation and character design, the Microduck is adjacent rather than central — but the RL-on-consumer-hardware angle is worth tracking. Motion capture and reference generation for character animation increasingly pull from real-world movement data, and community-trained locomotion behaviors from an open platform like this could eventually feed back into the kinds of reference libraries that inform AI image generation and character rigging workflows.
At $399 with a pre-Christmas ship date, the Microduck is priced at impulse-buy territory for a serious hobbyist — about the cost of a mid-range GPU upgrade or a few months of a pro creative subscription. Pollen Robotics is betting that price point, combined with Hugging Face's existing community of ML practitioners, is enough to seed a real ecosystem of trained behaviors. If it works, the more interesting robots won't be the ones Pollen ships — they'll be the ones the community trains.
