Pollen Robotics

#1By August 2026, the robotics race was getting bigger in almost every sense. Bigger funding rounds. Bigger machines. Bigger promises about humanoids that might one day do all kinds of human work.

#2Then Pollen Robotics, a small French robotics company, unveiled a duck.

#3About 25 centimeters tall. Roughly 800 grams. Two legs, a beak, one big eye. Roller skates, if you wanted them.

#4Microduck.

#5$399.

#6But the oddest thing wasn't the duck part. It was what Pollen chose not to finish.

#7Microduck came with behaviors of its own, but Pollen also gave people a simulator and training tools so they could teach it new movements and move those learned behaviors onto the real robot.

#8Most robotics companies were trying to answer, "How much can we make the robot do?"

#9Pollen was also asking, "What happens if other people decide what it does next?"

#10That question had been following the company for years—long before there was a Microduck, or even a Pollen Robotics.

#11Matthieu Lapeyre was doing AI research at Inria in France. His team, FLOWERS, studied how artificial agents and robots could learn through interaction with people and the real world.

#12At some point, Matthieu wanted to spend less time only working on algorithms and more time building the machine those algorithms might actually move.

#13That was a meaningful detour for a researcher.

#14His supervisor, Pierre-Yves Oudeyer, let him take it.

#15So Matthieu started building a robot. Pierre Rouanet and others joined in, and the project became Poppy, a small humanoid whose body could largely be made with a 3D printer.

#16Then the team did something that turned out to matter even more than the robot's gait.

#17They opened it up.

#18Other researchers could inspect it. But so could teachers, makers, FabLabs, artists—people who weren't simply waiting for the Poppy team to tell them what the machine was for.

#19They changed it. Rebuilt parts of it. Tried things the original team hadn't planned.

#20Pollen didn't even exist yet, and its future founders were already watching something important happen:

#21They could make the robot. Other people could make what came next.

#22By 2015, the Poppy project was already talking about an ecosystem where communities could build compatible modules for one another.

#23It sounds like a neat first chapter now.

#24At the time, it wasn't the beginning of some carefully mapped road to a $399 duck.

#25There was no such road.

#26Pollen Robotics was founded in 2016.

#27The company didn't start by raising the kind of giant war chest that later became common in humanoid robotics. It mostly bootstrapped, selling robots and projects and using that money to keep building.

#28That made Pollen flexible.

#29It also made mistakes, delays, and cash problems much harder to absorb.

#30Even Reachy, the robot that eventually became synonymous with Pollen, wasn't the original plan for CES 2020.

#31The team had been working on an AI arcade machine.

#32Then, about two months before CES, they changed course.

#33They built Reachy instead.

#34Big eyes. Two arms. A robot made for interaction.

#35It worked well enough to become Pollen's flagship platform.

#36But a good robot doesn't automatically make a small hardware company safe.

#37A couple of years later, Pollen found out just how unsafe things could get.

#38The global component shortage hit in 2022.

#39Pollen needed parts to build Reachy.

#40Those parts weren't arriving.

#41No completed robot meant no delivery. No delivery meant no revenue.

#42For a company with hundreds of millions in the bank, that might have been a supply-chain headache.

#43For Pollen, it became a survival problem.

#44The company was already using a bank overdraft that let its account go down to €50,000 below zero.

#45It asked the bank to double that limit to €100,000.

#46The bank agreed.

#47A little more runway.

#48And right in the middle of this, Pollen was preparing for the ANA Avatar XPRIZE, a global competition for robotic avatars that could let a remote operator see, interact, and perform physical tasks through a machine.

#49Ninety-nine teams entered.

#50There was another problem: the existing Reachy wasn't strong enough for everything the finals demanded. Pollen needed tougher joints and a more capable machine.

#51During a component shortage.

#52Some parts took months to arrive. Spares were thin.

#53If something important failed, there might not be another one to swap in.

#54They went anyway.

#55Pollen finished second.

#56The prize was $2 million.

#57Not long before, the company had been asking its banker for another €50,000 of negative breathing room. Now it had enough money to keep going.

#58Matthieu later thanked the banker who had extended the overdraft when things were tight.

#59And the stronger Reachy they built for the competition didn't disappear after the final.

#60It evolved into Reachy 2.

#61Reachy 2 turned into a serious research platform. Reachy robots made their way into hundreds of deployments across more than 20 countries, including research settings at Cornell and Carnegie Mellon.

#62But Reachy 2 cost about $70,000.

#63You could open the software. You could make the platform hackable. But if the actual machine still lived mostly in funded labs, there was an obvious ceiling on who could participate.

#64Then Hugging Face moved deeper into robotics.

#65Its AI community already ran on a familiar loop: someone publishes a model, someone else downloads it, adapts it, builds on it, and shares again.

#66In 2024, Hugging Face launched LeRobot.

#67Pollen had spent years making robots other people could modify.

#68Hugging Face had spent years building infrastructure around AI that other people could modify.

#69The two started working together around Reachy 2.

#70Then, in April 2025, Hugging Face acquired Pollen Robotics.

#71It'd be easy to tell this as the moment Pollen "became open source."

#72That's backwards.

#73Pollen had been doing open robotics since Poppy.

#74What changed was the scale of the affordability question.

#75How cheap could a real robot get without becoming just a toy or an empty development kit?

#76A few months later, Pollen had an answer.

#77Reachy Mini wasn't Reachy 2 shrunk down.

#78It was Reachy 2 stripped down.

#79No arms. No hands. No legs. No object manipulation.

#80What stayed was what Pollen wanted for a small AI interaction platform: cameras, microphones, speakers, expressive head and body movement.

#81Starting price: $399.

#82Pollen described the goal in four words.

#83Affordable. Easy to use. Hackable. And cute.

#84That last one mattered in practice.

#85An early tester had a five-year-old daughter. She saw Reachy Mini and wanted to carry it around the house.

#86She couldn't.

#87A cable was holding it in place.

#88The team took that reaction seriously enough that it helped push them toward a Wireless version with a battery and onboard computer.

#89The engineers had been looking at an AI platform.

#90The kid was looking at something she wanted to take with her.

#91Then thousands of other people got their hands on one.

#92And they started adding things.

#93Reachy Mini grew past 10,000 units.

#94Pollen didn't try to remain the sole source of every new feature. It built a way for users to make and distribute apps for the robot.

#95An app store, basically.

#96People made games. Dance apps. Storytelling experiences. Telepresence tools.

#97The pattern was starting to rhyme with Poppy.

#98Back then, Pollen's future founders had released a robot and watched teachers, makers, and researchers turn it into things they hadn't predicted.

#99With Reachy Mini, they didn't just hope that would happen.

#100They designed a place for it to happen.

#101Then they started lowering another barrier: coding.

#102AI coding tools could help turn a description of an idea into a Reachy Mini app.

#103One of the people Hugging Face highlighted while showing this off was Joel Cohen.

#104He was 78.

#105So the potential builder was no longer just the robotics graduate student who knew the stack already.

#106Still, Reachy Mini had one limitation no app could solve.

#107It couldn't leave the desk.

#108While building Reachy Mini, the Pollen team kept coming back to a question.

#109What if the little robot's main job wasn't to sit on a desk and interact with you?

#110What if its job was to move through the world?

#111That question became Microduck.

#112Pollen brought back the thing Reachy Mini had sacrificed:

#113A body that could actually go somewhere.

#114But they didn't go back to a $70,000 machine.

#115About 25 centimeters. About 800 grams. Fifteen motors. $399.

#116Microduck could walk, get back up after a fall, kick a ball, pick things up with its beak, even move on roller skates.

#117And Pollen didn't want it to be one of those robotics kits that's technically exciting but only after you lose a weekend to setup.

#118It had to do something immediately.

#119So Microduck shipped with seven learned behaviors.

#120No code required just to make it move.

#121But if someone wanted to go deeper, that door was open too.

#122An SDK. A simulator. A reinforcement-learning stack.

#123Instead of scripting every joint angle by hand, reinforcement learning lets a robot try enormous numbers of movements in simulation and gradually learn a behavior.

#124Train something in the simulator, then move that learned policy onto the real Microduck.

#125A new dance. A different recovery motion. Some movement Pollen never shipped.

#126The idea wasn't, "We didn't finish the robot, so you do it."

#127It was closer to:

#128We finished enough for you to enjoy it. We just don't want to decide where its abilities stop.

#129At this point, Pollen can sound like a company whose answer to everything is "open source all of it."

#130Microduck isn't that simple.

#131The SDK is open. The simulation and reinforcement-learning tooling are open. The learned behaviors can be studied and retrained.

#132But the full mechanical and electronic design files aren't published.

#133Pollen itself has been clear that Microduck shouldn't be described as fully open-source hardware.

#134So the pattern isn't simply open everything.

#135It's more like opening the layer that lets someone else build the next thing.

#136With Poppy, people could alter the body. With Reachy, researchers could build experiments on the platform. With Reachy Mini, users could make apps.

#137With Microduck, the interesting layer is the process of teaching the body new physical behaviors.

#138And Pollen has said where it hopes that goes.

#139On Hugging Face, one person's AI model can become raw material for someone else's work.

#140Pollen wants physical behaviors to become easier to share in a similar way—along with the environments and training recipes needed to reproduce them.

#141Which leaves a much stranger possibility.

#142Microduck opened for preorders on August 27, 2026.

#143Orders came in fast. Within roughly a week, reports put the number above 10,000.

#144That doesn't mean the experiment has worked.

#145Most of those robots haven't reached users yet.

#146We don't know how many owners will actually train new behaviors. We don't know how many will share them. We don't know whether a real marketplace or community of physical skills will emerge.

#147But having thousands of identical robot bodies changes what's possible.

#148One person trains a dance.

#149They share the training setup.

#150Someone else takes it and runs it on their own Microduck.

#151The second robot may not have to learn the whole thing from scratch.

#152A skill that used to live inside one research robot could start moving between machines almost like software.

#153That's still a future Pollen has to prove.

#154But one of the people trying to prove it has already watched several versions of this company come and go.

#155Back in 2019, Augustin Crampette was an intern at Pollen Robotics.

#156He worked on a robot joint called Orbita, inspired by biological movement. That work later became part of the technology inside Reachy.

#157Then the company went through just about everything a small hardware company can go through.

#158Parts vanished. Deliveries stalled. The bank extended the overdraft. The team won $2 million at a robotics competition. Reachy 2 became a $70,000 research platform. Hugging Face bought the company. Reachy Mini reached more than 10,000 units.

#159Then Pollen published the names of the five people on Microduck's core team in 2026.

#160Augustin Crampette was there.

#161Seven years earlier, he was an intern working on one joint inside somebody else's robot.

#162Now he's helping build a tiny robot whose owners may end up teaching it movements its own creators never thought to make.

#163Hugging Face grew by making it easy for people to share, download, and build on AI models.

#164Then it launched LeRobot in 2024. A year later, it bought Pollen Robotics outright.

#165Why did a company built around sharing AI models decide it needed robot bodies too?