Hugging Face Microduck: The $399 Open-Source Robot Duck You Can Train

On August 27, 2026, Hugging Face unveiled Microduck: a 9.8-inch bipedal robot duck that waddles, kicks a ball, picks things up with its beak, roller-skates — and, more importantly, learns new behaviors through reinforcement learning. It's available for preorder at $399, with first deliveries targeted before Christmas 2026 in North America, Europe, and the UK.

It's a strange product from a company known for hosting AI models. And that's exactly the point.

The hardware

Microduck was developed by Pollen Robotics, the French robotics startup Hugging Face acquired in April 2025. Inside a 25cm, 800-gram body:

Each robot gets a permanent audio identity — its own creature-like voice, assigned at first boot and kept for life. It doesn't chat like an assistant; it quacks. The design is deliberately approachable without hiding the machinery.

One clarification that matters to developers: the software is open source (Apache-2.0), the hardware is not. The SDK, simulation environment, and full RL training stack are on GitHub; the mechanical and electronic design files are proprietary.

The real feature: sim-to-real reinforcement learning

What separates Microduck from a $399 desk toy is the training loop:

1. A behavior (a new walking gait, a recovery maneuver, a trick) is trained in a physics simulation using proximal policy optimization (PPO)
2. The trained policy is deployed directly to the robot — no hand-coded control
3. You can adjust the simulation, retrain, and redeploy — without repeatedly sending your actual duck tumbling across the room
4. Trained behaviors can be shared back to the community as downloadable policies, the same way people share AI models

The 7 factory moves were themselves trained in simulation this way. Hugging Face co-founder Thomas Wolf calls it "the first truly accessible reinforcement-learning robot." Sim-to-real has historically required thousands of dollars of lab hardware and months of tuning; Microduck compresses that entry price to $399 plus a laptop.

Why Hugging Face is building robots

Microduck isn't a side quest — it's the continuation of a deliberate strategy to bring open-source distribution to physical AI:

• 2024: Launched LeRobot — open models, datasets, and tools for robotics
• 2025: Acquired Pollen Robotics; released the SO-101 robotic arm; introduced HopeJR (full-size humanoid, 66 movements) and Reachy Mini (expressive desktop robot, 10,000+ units sold)
• 2026: Microduck — the first product aimed squarely at consumers and hobbyists

The company is explicitly applying its core playbook — hub + open weights + community sharing — to robot behaviors instead of language models. CEO Clem Delangue frames it as "the era of affordable open-source robots."

And the hardware is only the consumer end of a bigger bet. Microduck's "AI that acts" positioning (versus Reachy Mini's "AI that interacts") points at world-model training and physical AI — the frontier where NVIDIA is also spending heavily.

Context you should know before ordering

• Microduck won't do your chores. It's a developer platform wearing a duck costume. The honest test is how well cheap hardware bridges the sim-to-real gap on complex motions.
• Shipping target is "before Christmas 2026" — preorder language, first production run, treat the date as soft.
• Accessories cost extra: $39 pack (rollers, ball, laser pointer, NFC tags), $119 dev pack (spare motors, NFC tags), spare batteries and chargers sold separately.
• It's designed to fall. Pollen's own framing: the duck is "built to move and ready to fall" — a reminder that physical AI agents fail, and developers should design for that from day one.
• It lands mid-acquisition. Microduck went on preorder the same week reports surfaced that NVIDIA is buying Hugging Face for $12.9B — meaning the world's biggest GPU company is about to own a consumer robot line.
• Four colorways at launch: Cream, Graphite, Lavender, and Sky.

Bottom line

The Microduck is either the beginning of accessible physical AI or the most technically advanced way to kick a tiny ball across your living room. Both can be true. What's genuinely new is the price point on the sim-to-real loop: for $399, reinforcement learning on real hardware stops being a lab privilege. Pollen says the duck is "about 10x more fun in a group" — robot soccer leagues sound like a joke until you remember this is exactly how open-source communities bootstrap ecosystems.

Related: Hugging Face guide · NVIDIA buys Hugging Face · Perplexity Portable Computer · AI Agent Guide

Next Steps

Understand the company behind the duck and the AI techniques that train it.

how do AI agents work — return to the complete AI agent architecture guide.