Why a duck?
The honest answer is that the duck came second. Pollen Robotics set out to build the smallest serious platform for learning movement: a robot whose whole job is to leave the desk, explore, fall over and get back up. A two-legged gait at 25 cm tall waddles — so instead of disguising it, the team leaned in: rounded body, an articulated beak, big single eye, four friendly colors (Cream, Graphite, Lavender and Sky).
The character is not decoration, it is engineering psychology. A robot that learns by trial and error falls a lot. When a research humanoid falls, it looks broken. When this duck face-plants and then hauls itself upright — one of its trained behaviors, not a scripted trick — it looks like learning. That difference is exactly why a machine like this can live in a family home or a classroom instead of only in a lab.
What Microduck can do out of the box
Microduck arrives with seven trained behaviors that work the moment you switch it on (a gamepad is included in the box):
- Walk — a velocity-tracking gait you steer with the gamepad
- Sit and stand — crouch down and get back up on command
- Kick — yes, it kicks a small ball
- Grab — its beak is a working gripper for light objects (around 100 g: markers, socks, small toys)
- Roller-skate — with the skate attachments from the accessory pack
- Get back up — self-recovery from common fall positions
- Quack — every unit generates its own unique voice the first time it wakes up, and keeps it for life
Behind those seven sits a published registry of 13 trainable tasks. The seven in the box are the starting point, not the ceiling — and that is the entire point of this robot, as the next section explains.
The real story: a robot that learns (sim-to-real)
Almost every robot you can buy under $500 executes programs a human wrote line by line. Microduck does something fundamentally different, the same technique used in research labs and by companies like Google DeepMind:
- Train in simulation. Each behavior starts life in a physics simulator (MuJoCo, via the mjlab framework), where thousands of virtual Microducks — 4,096 parallel environments — learn by trial and error using an algorithm called PPO. A walking policy can live a lifetime of falls in about an hour or two of computing.
- Harden it. A technique called domain randomization constantly varies the simulated floors, friction and physics, so the policy learns to cope with a world that never matches the simulation exactly.
- Deploy to the real robot. The trained policy is exported to ONNX and runs directly on the robot at 50 Hz — fifty decisions per second, on board, no cloud.
- Watch reality argue back — and retrain. When the real carpet defeats the simulated floor, you adjust and train again. That loop has a name: sim-to-real. It used to require a research budget. Microduck puts the whole loop on your desk for $399.
We are publishing a plain-English, step-by-step guide to this loop — What is sim-to-real? — in our Guides section in November.
Open source — precisely
You will see Microduck described elsewhere as "100% open source." Here is the accurate version:
- The software is open source under the Apache-2.0 license: the robot's control stack, the Python SDK, the simulation environments and the complete reinforcement-learning training pipeline are published on GitHub by Pollen Robotics.
- You can download policies other owners have trained, publish your own on Hugging Face, and inspect every reward function behind the behaviors in the box.
- The hardware is not fully open source. The electronics and mechanics are a commercial product, not an open-hardware design you can fabricate yourself.
We think that distinction matters, and we'd rather be the site that states it plainly.
Who Microduck is for — and who should skip it
Microduck is for you if you are a developer or student who wants to learn reinforcement learning on real hardware; a maker who has trained models in simulation and wants to feel the sim-to-real gap firsthand; or a family with a technical parent who wants a robot a child can teach, not just remote-control.
Skip it (for now) if you want a conversational companion that talks back — that's Reachy Mini's job, see our comparison: Microduck vs Reachy Mini. Skip it if you expect a household helper: at 800 g with a 100 g gripper, this is a learning platform, not a servant. And skip it if you want a finished toy for a young child to use alone — the play value for kids comes with an adult in the loop.
Price, regions and shipping status
Status as of October 2026: Pre-orders opened on 27 August 2026 at $399, before taxes and shipping. First deliveries are targeted before Christmas 2026 in North America, Europe and the UK, with Norway, Switzerland, Japan and South Korea also listed at launch. EU buyers should budget for VAT on top of the US list price.
We update this section whenever Pollen Robotics publishes new shipping information — the changelog is at the bottom of this page.
Our take: why one is coming to our lab
We are a family in Germany's Black Forest, and we have a Microduck on pre-order. Not as a gadget — as the teaching machine this site exists to document. Our 9-year-old is learning logic and programming; our 5-year-old learns by watching things try, fall and try again. So do neural networks, it turns out.
The moment our unit arrives, this page stops being a summary of other people's announcements: unboxing, first boot, first falls and our first self-trained policy all land in the lab log, with video on our YouTube channel. Our honest 30-day review will be added to this page once we've actually lived with the robot — we'd rather earn the word "review" than borrow it.