Signals Inbox·August 30, 2026·Robotics

Why are so many people buying Microducks?

Microduck is selling because it makes physical AI unusually accessible: $399 buys developers a real reinforcement-learning robot they can hack, train and keep next to a laptop, while its personality and viral design pull in people who would normally never shop for robotics hardware.

Q1Are people really buying that many Microducks?

Yes, Microduck demand is currently much bigger than a normal hobby-robot launch: Hugging Face says buyers ordered more than $2.6 million worth in the first 24 hours, and new customers are already facing a long delivery backlog.

Thomas Wolf, Hugging Face's co-founder and chief science officer, said orders crossed $1 million only hours after the launch. By the end of the first 24 hours, the company was above $2.6 million. Resell Calendar, which tracked the launch from a reseller angle, reported more than 5,000 sales over that period.

The $399 base price helps put the number in perspective. Even if every dollar had gone toward the robot itself, $2.6 million would equal about 6,500 Microducks. The real unit count should be somewhat lower because buyers can add a $39 charger pack, a $39 accessory pack or a $119 developer pack. Either way, we are talking about thousands of physical robots ordered almost immediately.

The backlog makes the demand harder to dismiss as a flashy revenue screenshot. Business Insider reported that new Microduck orders now face roughly four to six months of waiting after demand ran ahead of Pollen Robotics' initial production plans.

Q2Did Microduck demand disappear after the launch rush?

No, Microduck sales slowed after the first few hours, but demand stayed unusually high through the rest of the first day.

The first six hours generated more than $1 million, or roughly $167,000 per hour. The following 18 hours added more than $1.6 million, which works out to about $89,000 per hour. The pace therefore fell by roughly half after the earliest buyers rushed in, but it did not collapse.

For comparison, Petoi's Bittle robot dog raised $567,218 from 2,052 Kickstarter backers over a 44-day campaign. Microduck generated roughly 4.6 times that amount in its opening day. Kickstarter and direct preorders work differently, but the gap is too large to explain away as normal launch excitement.

Thomas Wolf also said at one point during the launch that Microduck was selling at roughly one unit every five seconds. We should treat that as a snapshot during the rush rather than a sustainable sales rate, but it shows how intense the initial buying became.

There was clearly a launch spike. There was also enough demand behind it to keep orders coming at a very high rate once the first wave had passed.

Q3Who is buying Microducks right now?

Right now, Microduck looks much more like a hit among developers, AI enthusiasts and robotics hobbyists than a mass-market toy bought mainly by families.

Pollen Robotics describes Microduck as a robot for physical AI and reinforcement learning. Its store already has a dedicated option for organizations ordering ten or more units, with bulk pricing and import support. The $119 developer pack includes spare motors, motor cables, extra batteries, tools and Hugging Face compute credit. Those are strange accessories for a conventional electronic toy and very normal ones for hardware people expect to experiment on.

The conversation around Microduck points in the same direction. In the robotics and reinforcement-learning communities on Reddit, people are discussing PPO training, MuJoCo Warp, sim-to-real transfer, autonomous navigation, SLAM and grasping. A Pollen engineer joined those discussions to explain how the robot is trained and what developers could build next.

There is clearly a second group attracted by how Microduck looks and behaves. The launch spread well beyond specialist robotics circles, and videos of the duck walking or skating reached mainstream technology audiences quickly.

We still do not have a customer breakdown from Pollen showing how many buyers are developers, universities, parents, collectors or resellers. For now, though, the product itself and the early conversations around it strongly point toward a technically skewed first wave.

Q4Why does a $399 Microduck feel cheap to robotics people?

A $399 Microduck sits in a sweet spot where the robot is still an expensive gadget for a casual buyer but surprisingly affordable for someone who wants to experiment with physical AI.

Petoi's Bittle X robot dog currently starts around $319. Unitree's much more capable Go2 quadruped starts at $1,600. Unitree's R1 humanoid starts around $4,900, while its G1 starts at $13,500. Research robots can quickly move into five-figure or even six-figure territory.

Microduck gives an individual developer 15 motors, onboard computing, a camera, LiDAR, two IMUs, microphones, a speaker and a working reinforcement-learning stack for roughly the price of a high-end consumer gadget.

The comparison with Bittle is especially useful. Spending around $300 on a programmable small robot already has an established market: Petoi says it has shipped more than 30,000 robots. Microduck asks only about $80 more than the current Bittle X starting price while pushing much harder into reinforcement learning and sim-to-real development.

That price changes who can realistically experiment. A developer can buy a Microduck personally, leave it next to a laptop and start breaking things without needing a university lab or a company robotics budget.

Q5What can a Microduck actually do?

Microduck can currently walk, sit, crouch, kick a ball, pick up small objects with its beak, recover after falling and move on rollers, while developers can train completely different behaviors themselves.

The robot is about 25 centimeters tall and weighs around 800 grams. Inside are 15 motors, a wide-angle camera, an 8×8 time-of-flight LiDAR sensor, two IMUs, microphones, a speaker, NFC, Wi-Fi and Bluetooth. A Rockchip RK3566 handles onboard computing, and the control policies run at 50 Hz.

Pollen ships seven trained movements with the robot, so someone can use it immediately with the included game controller. The deeper part starts when the owner wants the robot to behave differently. The software, simulator and reinforcement-learning training stack are available publicly, and trained policies can be exported to ONNX and run on the physical Microduck.

The articulated beak is also more important than it first looks. Microduck has no arms, so Pollen gave the head a second job: the robot can lower itself, grasp a small object with its beak and stand back up. That gives developers at least a basic manipulation problem to work on rather than limiting everything to locomotion.

Microduck is still a very small machine with very limited physical reach. The impressive part is how much robotics experimentation Pollen has squeezed into that size.

Q6Is Microduck really an AI robot?

Yes, Microduck genuinely uses machine learning to control its body, although buyers should not expect the kind of autonomous AI companion people imagine when they hear the phrase "AI robot."

Microduck's movements rely on neural policies trained through reinforcement learning. Pollen's public Microduck RL repository shows developers how to train those policies with PPO inside MuJoCo, add domain randomization and actuator modelling, export the resulting model to ONNX and run it on the real hardware.

That is real machine learning controlling a physical robot. The software stack currently includes tasks for walking, standing, recovering from falls, sitting, touching the ground with its beak and kicking a ball.

Autonomy is much less developed. A Pollen engineer described navigation, SLAM, natural-language control and better grasping as areas still open for developers to explore. The company's public roadmap also shows that the autonomous "brain" remains unfinished.

So when Pollen calls Microduck a physical-AI platform, the description holds up. Calling it an intelligent home companion today would be much harder to defend.

Q7Why are developers excited about training Microduck in simulation?

Developers like Microduck's simulator because they can let a virtual duck fall over thousands of times before risking the real robot.

Training physical movement through reinforcement learning involves enormous amounts of trial and error. Doing every bad attempt on hardware wastes time, requires someone to reset the machine and puts motors and mechanical parts through unnecessary stress.

Microduck's workflow moves most of that pain into MuJoCo. A developer can train thousands of simulated versions in parallel, adjust rewards and physics parameters, test the result virtually and then export the trained policy onto the physical robot.

Pollen's own public training instructions show how far this has been pushed toward ordinary developers. The Microduck RL repository describes a usable walking policy as taking roughly one to two hours to train across 4,096 simulated environments on suitable GPU hardware. Developers without their own GPU can also send jobs to Hugging Face's compute infrastructure.

The small physical robot helps once a policy reaches the real world. An 800-gram Microduck falling onto the floor creates a very different problem from a large humanoid falling in a lab.

This combination is probably one of the strongest technical reasons people are buying Microducks now: developers get to work on a real sim-to-real problem without making every mistake on an expensive machine.

Q8Does open source really make Microduck more valuable?

Yes, open source gives Microduck a much better chance of becoming more useful after purchase because owners are able to change the software, retrain its movements and share what they build.

Pollen has published the robot runtime, reinforcement-learning environments, simulator and training tools. Developers can inspect the policies controlling the robot rather than treating its behavior as a black box.

We have already seen this model work with Pollen's previous robot. Hugging Face says more than 10,000 Reachy Minis have reached users around the world. Its community has published more than 200 apps created by over 150 people, and Hugging Face says most of those creators had never programmed a robot before.

That history makes the Microduck bet easier to understand. Pollen already knows that giving people an inexpensive robot plus open software can produce applications the original team never planned.

Microduck pushes the same idea into movement. Instead of sharing a desktop-robot application that plays chess or acts as a language tutor, developers can eventually share a walking style, a recovery move, a dance or some completely new physical behavior.

Whether sharing Microduck policies becomes genuinely easy is still being worked out. The company's public roadmap currently includes a dedicated model channel for installing and rolling back individual policies. The distribution layer is young, but the direction is pretty obvious.

Q9Did Microduck suddenly come out of nowhere?

No, Microduck looks sudden because the commercial launch went viral, but the robot sits on top of several years of open robotics work and a surprisingly large existing community.

Pollen Robotics has been building open robots since 2016. Hugging Face acquired the French company in 2025 as it expanded its own work around open-source robotics and LeRobot.

More specifically, Microduck grew out of the Open Duck Mini project, an open attempt to build a small BDX-style droid. That project already attracted roughly 3,400 GitHub stars and more than 400 forks, with Hugging Face and Pollen sponsoring the work.

Then came Reachy Mini. Pollen says more than 10,000 of those desktop robots are now in circulation, giving the company a ready-made group of developers who already understand its approach to open robotics.

The Microduck launch therefore looks a lot less random than it did on day one. Pollen had already tested the hardware community, the open-source model and demand for affordable robots separately. Microduck brought those pieces together in a much more visually appealing product.

Q10Is Microduck selling because it looks cute?

Yes, Microduck's appearance is probably doing a huge amount of the selling, and the clever part is that the cute features also show what the robot can technically do.

A 25-centimeter duck with one eye, little legs and a strange voice is much easier to care about than a list of reinforcement-learning specifications. Pollen offers Cream, Graphite, Lavender and Sky versions and shows the robot on bedroom floors and desks rather than presenting it like industrial equipment.

The rollers are a good example. They cost only $39 as part of an accessory pack, yet skating quickly became one of the most widely shared demonstrations of Microduck. The Verge, TechRadar and other mainstream outlets led with the roller-skating duck rather than its PPO implementation.

That spectacle still tells us something useful about the machine. Keeping a small biped balanced while it moves on rollers is a locomotion problem, exactly the kind of robotics Pollen wants developers to experiment with.

The robot also recovers from falls, kicks a ball and grabs objects with its beak. All of those movements work visually without someone first understanding reinforcement learning.

Pollen has made the technical demo entertaining enough that millions of people can understand it in a few seconds. That is rare in developer hardware, and it helps explain why Microduck escaped the normal robotics audience so quickly.

Q11Is Hugging Face the real reason Microduck took off so fast?

Hugging Face gave Microduck a huge head start by putting an unfamiliar robot in front of an existing global AI developer community that already trusts the company's open-source ecosystem.

An unknown robotics startup selling a new biped has to solve several problems at once. It needs people to discover the product, believe the hardware will ship, learn an unfamiliar software stack and hope other developers eventually show up.

Microduck launched with much of that already solved. Hugging Face hosts millions of AI models and is used by more than 50,000 organizations. Its robotics ecosystem already includes LeRobot, Pollen Robotics, Reachy Mini, simulators and community applications.

As seen above, Reachy Mini has already put more than 10,000 Pollen robots into users' hands. That gives Microduck a natural audience of people who understand what Hugging Face is trying to build around open robotics.

AI developers are also exactly the people most likely to understand why a $399 reinforcement-learning robot is interesting. Hugging Face did not have to explain physical AI to the entire consumer market before finding its first few thousand customers.

We cannot separate the brand effect from the product effect with the data available today. Still, it is difficult to imagine an identical robot from an unknown company producing the same opening rush.

Q12Are resellers making Microduck look more popular than it really is?

Resellers are adding some extra Microduck demand now, but they appear to have joined the rush after genuine buyer demand had already created scarcity.

Once the delivery window stretched into several months, reseller communities noticed. Resell Calendar explicitly told its audience that early Microduck orders could become profitable flips if buyers later wanted the robot without waiting.

That can reinforce the shortage. Resellers buy units because availability is tight, and those orders make availability even tighter.

The timing is useful here. Microduck had already crossed $1 million of orders within its first few hours, before a mature resale market had developed around the product. The reseller story gained strength after the backlog became obvious.

We have no public data showing what percentage of current orders belong to resellers, so putting a precise number on the effect would be guesswork. Speculative buying now exists and is probably making an already successful launch look slightly hotter.

It did not create the original Microduck rush.

Q13Are people already teaching Microduck new tricks?

Yes, developers are already publishing new Microduck behaviors, which is one of the most encouraging fresh signs that the product may become more than a one-week viral gadget.

The official robot comes with seven trained movements, but community experiments are already moving outside that list. On Hugging Face, developers have published Microduck policies for behaviors including a moonwalk-style backward gait, a polite bow and a flamingo-like single-leg balancing cycle.

These are still early experiments, and some have only been validated in simulation so far. A policy looking good in MuJoCo does not guarantee it will behave safely or cleanly on the physical robot.

The pace is still interesting. The robot has only just launched, yet people are already treating movement as something they can remix and publish rather than waiting for Pollen to release official updates.

The reinforcement-learning repository is moving too. Developers have opened discussions around alternative GPU support and walking policies, while the main Microduck software repository is actively documenting control, updates, recovery and future policy distribution.

The community is still tiny and young, so this evidence deserves less weight than the sales numbers. But this is exactly the behavior we would want to see if Microduck is going to become an open robotics platform.

Q14Is Microduck actually useful enough to be worth buying?

For someone who wants to learn robotics or experiment with physical AI, Microduck already has a convincing use case; for someone who wants a useful household robot, it currently offers very little.

A Microduck cannot clean a room, carry groceries, unload a dishwasher or perform meaningful household labor. Its body is only 25 centimeters tall, and its one articulated beak severely limits what it can manipulate.

The developer use case is much stronger. For a few hundred dollars, someone can experiment with locomotion, reinforcement learning, perception, sim-to-real transfer, fall recovery and basic manipulation on actual hardware.

That makes the value highly dependent on the buyer. Someone who simply wants a cute electronic pet may get bored after trying the seven included behaviors. Someone who wants to understand how a learned policy behaves differently on a physical floor than in a simulator can spend months working on the same machine.

The $119 developer pack is revealing here. Pollen includes three spare motors, spare motor cables, tools, batteries and compute credit. The company clearly expects some owners to push the hardware hard enough that maintenance becomes part of the experience.

Microduck makes the most sense today as a robotics playground that happens to be charming enough to sit in a living room.

Q15Could Microduck become a mainstream robot pet?

Microduck could become a much bigger consumer product, but today's buying rush still falls short of proving that ordinary households want an open-source robot pet at scale.

Clem Delangue has already set an ambitious target. He told Axios that he sees Microduck potentially beating the roughly 20,000-unit level reached by today's best-selling robots and said 50,000 units would be a great success. At the introductory price, that would represent roughly $20 million of robot sales.

There is evidence that the broader category can support tens of thousands of units. Petoi says it has shipped more than 30,000 programmable robot pets across more than 60 countries.

Microduck also has several advantages if Pollen wants to reach beyond developers. It works with a controller out of the box, has a recognizable personality, makes sounds, comes in different colors and can perform movements that people immediately understand.

The obstacle is depth. A mainstream robot pet needs to remain interesting without requiring its owner to open GitHub, train a policy or debug software. Microduck currently gets much of its long-term value from exactly those activities.

If the developer community starts producing behaviors ordinary owners can install in a few clicks, that equation could change quickly.

Q16Will people still care about Microduck after the novelty wears off?

We still do not know whether Microduck will become something people use for years, and this is now the most important question after the spectacular launch.

Seven built-in movements can generate great videos, but they are unlikely to sustain months of engagement by themselves. The roughly one-hour battery life also makes Microduck better suited to sessions and experiments than to roaming around a home all day.

The stronger case for longevity comes from software. Developers are already publishing new movements, and Pollen is building the infrastructure needed to make policies easier to distribute. Reachy Mini previously showed that a relatively small robot community can produce hundreds of applications.

Microduck's current software roadmap also tells us how early this still is. Pollen lists important pieces such as the dedicated app experience, easier model distribution and the autonomous brain among the work that remains.

The real test is simple. If owners are still sharing new movements and useful behaviors once thousands of physical Microducks have actually arrived, the launch will look like the beginning of an ecosystem. If most activity remains concentrated around skating videos and the original seven tricks, Microduck will look much more like an exceptionally successful gadget launch.

For now, the first community experiments are encouraging, but the physical robots have barely had time to reach users. Confidence should be lower here than on the question of whether initial demand is real.

Q17Why are so many people buying Microducks?

People are buying Microducks because Pollen Robotics has managed to make serious physical-AI experimentation cheap, approachable and unusually fun at exactly the moment more AI developers want to move beyond software.

The demand itself is real. Thousands of robots were ordered almost immediately, the first-day order value reached several million dollars, and production quickly developed a multi-month backlog. Resellers are adding some heat now, but the rush started before the resale story.

Price explains another large part of it. Robotics enthusiasts can already buy programmable pets for around $300, while more capable legged robots quickly cost four figures and humanoids cost far more. Microduck lands close to the low end of that range while giving developers a working reinforcement-learning and sim-to-real environment.

Then Pollen made the robot enjoyable to watch. A duck that falls over, stands up, grabs something with its beak and roller-skates sells the idea of physical AI far better than a technical specification sheet. Hugging Face put those demonstrations directly in front of one of the world's largest communities of AI builders.

The most interesting development is what happened immediately after the launch. Developers have already started training different movements and publishing them. That is the part worth watching now, because it could turn a funny $399 robot into a genuine platform for experimenting with physical AI.

So our answer is clear: people are buying Microducks because the product combines a rare set of things that normally live in separate markets. It is technically interesting enough for robotics developers, affordable enough for individuals, open enough to hack, simple enough to play with immediately and cute enough that people want one before they have even decided what they will build with it.

The buying craze is real. Whether Microduck becomes a lasting robotics platform is still open, but the launch has already shown something important: there are thousands of people willing to buy physical-AI hardware once it becomes accessible enough to keep next to a laptop.