Microduck: a $399 open-source physical-AI dev kit

Hugging Face opened pre-orders on August 27 for Microduck, a $399 bipedal robot duck. The price is the headline, but it is not the interesting part. The interesting part is that the entire reinforcement learning stack ships under Apache-2.0 — simulator, training code, and all seven shipped policies, retrainable.

What actually happened

Per TechCrunch and Pollen Robotics, Microduck stands 25cm tall, weighs about 800g, and runs 15 motors, a camera, a lidar sensor, two IMUs, and an articulated beak that can pick things up. The onboard policy loop runs at 50Hz. It ships with seven trained behaviors — walk, sit and stand, kick, grab, roller skate, and get back up after a fall — and every one of those policies is published rather than shipped as a black box.

The SDK, a MuJoCo simulation environment, and the full RL training stack are on GitHub under pollen-robotics/microduck. Base price is $399 before tax and shipping at introductory pricing, with a $39 Accessory Pack, a $39 Charger Pack, and a $119 Dev Pack. Pre-orders ship before Christmas. Hugging Face acquired Pollen Robotics in April 2025; Microduck sits below the Reachy Mini line in the same catalog.

CEO Clem Delangue described it as an open-source robot you can teach new tricks with reinforcement learning, positioning it in "the era of open-source affordable robots to democratize physical AI."

Why open-source robotics matters for your business

You are not buying a robot duck for the warehouse. But if physical automation is anywhere on your three-year map — pick-and-place, inventory scanning, shop-floor inspection, anything where a camera and an actuator replace a repetitive walk — the expensive part has never been the hardware. It is finding out whether the sim-to-real loop works for your process before you sign an integration contract.

That evaluation used to cost a five-figure dev kit and a robotics hire. It now costs $399 and a MuJoCo install. The specific thing worth testing is the loop itself: define a behavior, train it in simulation, deploy it to the physical unit, watch where it fails, retrain. If that cycle runs fast and honest on a duck, it will run slow and honest on a warehouse arm — same math, bigger budget. If your team cannot get the cycle working at all on a $399 platform with published policies and an open simulator, that is the cheapest possible answer to a very expensive question.

The pattern here is the one we keep flagging: open weights and open training code make a vendor's claims checkable. A robot that ships its policies is a robot whose behavior you can inspect, fork, and audit. A robot that ships a firmware blob is a subscription with legs. The argument we make about model vendors applies to anything with motors.

Sources: TechCrunch, Pollen Robotics.