Four robot ducks on a black backdrop, 25 centimetres tall, in cream and graphite and lavender and sky. That launch shot is doing a lot of work. Pollen Robotics, the French outfit Hugging Face bought last year, opened Microduck pre-orders on 27 August at 399 dollars, and the machine under the cuteness is real: 15 motors, a camera, an 8x8 time-of-flight lidar, a beak that grabs, seven reinforcement-learning policies you can retrain yourself. Where we'd slow down is the phrase everyone reached for. The software stack genuinely is Apache 2.0, runtime and training pipeline both. The mechanical and electronic design files are not published. So Microduck is an open software robot, not open source hardware, which is a step back from Reachy Mini. Worth knowing before you buy one as a teaching platform.
The short answer
Pollen Robotics, owned by Hugging Face, is selling a 25 cm walking robot for 399 dollars with 15 motors, a lidar, a grasping beak and seven retrainable RL policies. The runtime and the training stack are Apache 2.0 on GitHub. The mechanical and electronic design files are not, so the “open source robot” headline needs an asterisk. Pre-orders opened 27 August. The Christmas delivery target is already gone for new orders.
What Pollen actually shipped
Start with the hardware, because it’s better than the price suggests.
Twenty-five centimetres, about 780 grams, and 15 motors spread across the legs, neck and head. There’s an articulated beak that grabs things, which is a genuinely unusual choice: most robots this size get a fixed shell and a wiggle. Sensing runs to a front camera, an 8x8 time-of-flight lidar, a pair of IMUs, microphones and a speaker. Two NFC antennas, so the thing can read tags off the floor. Wi-Fi and Bluetooth.
That lidar spec deserves a beat, because “lidar” in a headline conjures a spinning drum on a robotaxi roof. An 8x8 ToF array is a 64-pixel depth sensor, roughly the class of part in a modern phone’s autofocus. It’s plenty for “is there a wall” and useless for mapping a room in detail. Not a criticism, just a right-sizing.
Image: Pollen Robotics (the four Microduck colourways: Sky, Graphite, Cream, Lavender)
The brain is a Rockchip RK3566 with an AI accelerator, 1 GB of RAM, 32 GB of storage, running a 50 Hz policy loop. That’s a cheap Android TV box chip, and it tells you exactly what this robot is for. Nothing resembling a language model runs on it. The onboard job is loading a small exported control policy and hitting its deadline fifty times a second, which the RK3566 does fine.
Power comes from a removable NP-F550. That’s the standard Sony-style camera battery you can buy anywhere for pocket change, and picking a commodity cell over a glued-in custom pack is the single most developer-friendly decision on the whole spec sheet. It’s good for about an hour. Buy three.
The licence, read carefully
Here’s the part the coverage skated past.
Two repositories are public and Apache 2.0: pollen-robotics/microduck, the runtime that manages the servo drivers and the control loop on the RK3566, and pollen-robotics/microduck_rl, the training side. The training repo is the interesting one. It’s not a black box you poke with an SDK. The reward functions are readable, the sim-to-real recipe is documented, and policies train in MuJoCo with PPO before exporting to ONNX for the robot to load. Pollen puts a gait at roughly one to two hours of training at 4096 parallel environments, which is an afternoon on a rented GPU rather than a research project.
So the RL claim is real. You can genuinely retrain all seven shipped behaviours, and the seven are a decent spread: walking, sitting, crouching, kicking, roller-skating, plus recovery from common falls.
What isn’t public is the mechanical and electronic design. No CAD, no schematics, no board files. Which means you can fork the brain and not the body.
I keep coming back to Reachy Mini on this. Pollen’s previous desktop robot was open on both sides, hardware and software, and that was the headline feature. Microduck quietly drops half of it. Nobody’s obliged to open a mechanical design, and manufacturing a walking biped at 399 dollars probably depends on nobody else printing it. But if you’re a university buying twenty units because the marketing said open source, check which half you’re getting.
The price, and the other price
399 dollars, 340 euros, introductory, before tax and shipping. That’s the number in every headline and it’s honest as far as it goes.
Then you remember it walks. Motors in a legged robot are consumables, and Pollen knows it, which is why there’s a 119 dollar developer pack holding three spare motors, five motor cables, two batteries, a charger, ten NFC tags and tools. There’s a 39 dollar charger pack with a dual charger and two spare cells. Neither is optional in practice if the robot is going to be used rather than displayed.
Call it 557 dollars for a kit that survives a semester. Still cheap for what it is. Just not 399.
The date already moved
This is the freshest thing in the story and it’s sitting on Pollen’s own store page rather than in any press release. The launch messaging said first deliveries were targeted before Christmas 2026. The product page now carries a notice saying the community ordered a lot of ducks, that Christmas can no longer be promised for new orders, and that the estimate is four to six months while production ramps.
Four to six months from late August lands somewhere in the first quarter of 2027. If you’re planning a course or a demo around this, plan around that number, not the one in the launch coverage. Honestly, a pre-order slipping under demand is the good version of this problem, and it beats the alternative where nobody ordered. But the gap between announcement and store page opened in under 48 hours, which is fast even by pre-order standards.
Who this is for
If you teach robotics or reinforcement learning, this is the most interesting thing in its price bracket, and the open training repo is why. A student can change a reward function, train a gait in MuJoCo over lunch, and watch the real robot walk differently. That loop, at 399 dollars, didn’t exist last year.
If you were hoping to design your own body around the electronics, you can’t, and the licence won’t get you there.
And if you’re watching this as a business signal rather than a purchase, it’s worth reading next to the reported Nvidia move on Hugging Face. A hub known for hosting weights now ships consumer hardware with a training stack attached, which is a different company than it was two years ago. The robot brains are getting cheaper and more open in parallel, as Google’s Gemini Robotics 2 benchmarks showed from the model side. Microduck is the cheap body arriving to meet them.
Order one if you’ll actually train something on it. Otherwise wait for the second batch, when the ship date means something.
Sources: pricing, colourways, launch regions, the full spec list and the revised delivery estimate come from Pollen Robotics’ own store listing and the Microduck product page, which also lists the accessory pack contents and prices. The runtime and training repositories are pollen-robotics/microduck on GitHub under Apache 2.0. Independent spec confirmation and the pre-order date come from Engadget and MarkTechPost, the latter also reporting the MuJoCo and PPO training figures and noting that the mechanical and electronic design files are not open. Reachy Mini prices are the ones Hugging Face published at that product’s launch.
Frequently asked questions
How much does Microduck cost and when does it ship?
Microduck is 399 dollars, or 340 euros, as an introductory pre-order price before tax and shipping. Pre-orders opened on 27 August 2026. The launch messaging targeted first deliveries before Christmas 2026, but Pollen's own store has already walked that back: it now says it cannot promise Christmas delivery for new orders and quotes an estimated four to six month wait while it ramps production. Launch regions are the US, Canada, the EU, the UK, Norway, Switzerland, Japan and South Korea.
Is Microduck actually open source?
The software is. Both the runtime repository (pollen-robotics/microduck) and the reinforcement-learning training repository (pollen-robotics/microduck_rl) are on GitHub under Apache 2.0, including the reward functions and the sim-to-real recipe. The mechanical and electronic design files are not published. That makes it an open software robot rather than open source hardware. Reachy Mini, the earlier Pollen product, was open on both sides, so this is a narrowing rather than an expansion.
What are Microduck's specs?
It stands 25 cm and weighs about 780 g. Fifteen motors are spread across the legs, neck and head, plus an articulated beak that can grasp small objects. Sensing is a front camera, an 8x8 time-of-flight lidar, two IMUs, microphones and a speaker, with dual NFC antennas. The brain is a Rockchip RK3566 with an AI accelerator, 1 GB of RAM and 32 GB of storage, running a 50 Hz onboard policy loop over Wi-Fi and Bluetooth. Power is a removable NP-F550 battery good for roughly an hour.
Can I train my own behaviours on it?
Yes, and that is the actual product. Seven policies ship pre-trained, covering walking, sitting, crouching, kicking, roller-skating and getting back up after a fall. All seven are retrainable. Training happens in MuJoCo with PPO across thousands of parallel environments, and the resulting policy exports to ONNX for the onboard loop to load. Pollen reports a gait converging in roughly one to two hours at 4096 environments, which puts a full training cycle inside an afternoon on a rented GPU.
Does Microduck run an LLM on board?
No. One gigabyte of RAM on an RK3566 is nowhere near enough, and Pollen does not claim otherwise. The onboard compute exists to run small exported control policies at 50 Hz, not language models. If you want an LLM in the loop, it runs off the robot and talks to it over Wi-Fi. Microduck also does not speak: it communicates through audio tones rather than speech, and each unit gets a persistent audio identity after setup.
