“Pay attention. Stay focused. Come here.”
That is not a teacher on a school trip. It is what Jensen Huang, the CEO of NVIDIA, said on stage to a robot shorter than a school ruler that had just tripped and wandered away from its friends.
Welcome to the funniest moment of GTC Berlin 2026, and probably the most talked about tech demo of the week. Four tiny robots walked out, sat down, fell over, and somehow became the stars of a stage full of chips and servers. The whole Microduck demo lasted only a few minutes, but people have been replaying it ever since.
So what exactly is the NVIDIA Microduck robot?
Here is the short answer. It is a 25 cm (10 inch) walking robot built by Pollen Robotics, the robotics team owned by Hugging Face. NVIDIA did not build it. But NVIDIA put four of them on stage to show how robots learn to move in a simulator first and then walk in the real world.
In this post, I will walk you through 7 facts about the Microduck. We will cover what it can do, what is inside it, how it was trained, how much it costs, and why a $399 duck robot matters to the whole AI industry.
Table of Contents
- Fact 1: Jensen Huang gave live orders to four tiny robots
- Fact 2: It is just 10 inches tall and packs 15 motors
- Fact 3: NVIDIA did not build it, so who did?
- Fact 4: It learned to walk inside a simulation
- Fact 5: The price, the sellout and the shipping date
- Fact 6: It is open source, and you can teach it new tricks
- Fact 7: Why this tiny robot matters for the future of AI
- Frequently Asked Questions
- Final Thoughts
Fact 1: Jensen Huang Gave Live Orders to Four Tiny Robots, and the Stumbles Made It Unforgettable
Most tech demos are polished. This one was not, and that is exactly why it worked.
What Actually Happened on Stage?
On Wednesday, October 21, four Microducks walked onto the stage by themselves. No one carried them. No one pushed them.
Jensen Huang greeted them and asked them to sit. They sat. He asked them to stand. They stood.
So far, so smooth.
Then he told them to take a nap.
Three of the four tipped straight onto the floor. The fourth kept walking for a few more steps, then fell onto its side. The crowd laughed, and so did Huang.
He encouraged the robots to get back up and then asked for somersaults. A couple of them pulled it off. A few stayed put. One needed extra encouragement.
And then came the best part. One robot wandered off, tripped, and picked itself back up. That is when Huang called it back with his now famous line. It walked over and sat down like a student who knew it was in trouble.
Why the Messy Demo Worked?
A perfect demo makes people suspicious. A messy one makes them lean in.
The falls told the audience something important. These robots were not playing a pre-recorded video. They were reacting live, and sometimes they got it wrong.
And look at what happened after every fall. The robots recovered. For a small legged robot, getting back up is one of the hardest skills to master, and the Microducks did it again and again.
Behind the jokes, NVIDIA had a serious message. According to Huang, the robots were trained in the cloud on large NVIDIA systems, then refined inside a digital twin running on an RTX Spark machine. We will break that down in Fact 4.
Related read: AI agents are learning to click buttons on a screen, while robots like the Microduck are learning to take steps in the real world. If you want the software side of that story, read our breakdown of how AI agents click buttons for you.
Fact 2: It Is Just 10 Inches Tall, Weighs Under Two Pounds and Packs 15 Motors
Do not let the cute face fool you. There is a lot of engineering inside this little body.
The Microduck specs show a robot that is small in size but serious in hardware. Here is the quick view.
| Feature | What You Get |
|---|---|
| Height | About 25 cm (10 inches) |
| Weight | Under 2 pounds |
| Motors | 15, spread across the legs, neck and head |
| Vision | Front camera plus a depth sensor |
| Balance sensors | Two small motion sensors (IMUs) |
| Gripper | An articulated beak that picks up light objects |
| Movement | Walks, sits, crouches and gets up after a fall |
| Extra | Optional wheels let it roller-skate |
| Built by | Pollen Robotics (a Hugging Face company) |
Let me explain why each of these matters.
Why 15 Motors Is a Big Deal?
Motors are the muscles of a robot. Fifteen motors across the legs, neck and head means the robot must coordinate many joints at the same time just to stay upright.
That is why a bad step or a small shove can send it to the floor. It is also why watching it recover feels so impressive.
What Those Two Motion Sensors Do?
An IMU, short for inertial measurement unit, works like the inner ear of a robot. It tells the robot when it is tilting, speeding up or about to lose its balance.
Without it, a walking robot is just a very expensive toy that falls over.
The Beak Is Not Just for Looks
The articulated beak can pick up relatively light objects from the floor. It is a clever, simple replacement for a hand.
It also shows that even a tiny robot can do more than walk and pose. It can interact with things around it.
And with wheels attached, the Microduck can roller-skate. Yes, really.
Related read: As robots get smaller, cheaper and smarter, not everyone is relaxed about where this is heading. We covered the warnings about AI superintelligence, robots and bio risk that experts have raised, and it is worth a read alongside this story.
Fact 3: NVIDIA Did Not Build the Microduck, So Why Was It on NVIDIA’s Stage?
Let’s clear up the biggest mix-up first.
Search for this robot online and you will see “NVIDIA Microduck” everywhere. It makes sense. The robots stood next to Jensen Huang, under NVIDIA’s lights, at NVIDIA’s event.
But the name is not quite right.
Who Really Built the Microduck?
The robot comes from Pollen Robotics, a team based in Bordeaux, France. Hugging Face bought the company in April 2025 to build cheap, open hardware.
Pollen opened Microduck pre-orders on August 27, 2026. That is almost two months before the GTC Berlin show. So the robot was already public before NVIDIA gave it a stage.
Where NVIDIA Fits In?
NVIDIA’s role is the tools and the platform. In NVIDIA’s GTC Berlin announcement, the company said Hugging Face and Pollen Robotics are using NVIDIA Isaac Lab, Isaac Sim and Newton to build Microduck. The goal is to help developers and students train, modify and share new physical behaviors.
So the fair description is this. Microduck is a Pollen Robotics and Hugging Face robot, built with the help of NVIDIA’s simulation software.
The $12.9 Billion Twist
Now the plot thickens. Look at how the dates line up.
| Date | What Happened |
|---|---|
| April 2025 | Hugging Face buys Pollen Robotics |
| August 27, 2026 | Pollen opens Microduck pre-orders at $399 |
| Late August 2026 | Reports say NVIDIA has agreed to buy Hugging Face |
| September 3, 2026 | NVIDIA confirms the deal |
| October 21, 2026 | Four Microducks walk onto the GTC Berlin stage |
NVIDIA confirmed the deal to buy Hugging Face for about $12.9 billion. It is the biggest outright purchase the company has ever made. NVIDIA also pledged to keep Hugging Face an open platform.
One important detail. NVIDIA expects the deal to close in the first half of 2027. That means it has not closed yet. Today, Pollen belongs to Hugging Face, not to NVIDIA.
My read is simple. If the deal goes through, Microduck’s maker would sit inside the NVIDIA family. That makes it easy to see why this tiny robot got such a big stage.
Related read: A $12.9 billion deal sounds huge, but it is only one number in a very big picture. If you want to know whether the AI boom can keep this pace, read our guide to the AI bubble numbers behind OpenAI, Anthropic and SpaceXAI.
The Bigger Plan: Be the Android of Robots
This also fits a pattern. When NVIDIA unveiled its robotics tools at CES 2026, TechCrunch described the strategy as an attempt to become the default platform for generalist robots, the way Android became the default for phones.
A cheap, friendly robot that anyone can train is a smart way to bring developers into that world.
Fact 4: The Duck Learned to Walk Inside a Computer Before Taking a Single Real Step
Imagine learning to ride a bike by falling off thousands of bikes at the same time, without a single scrape. That is close to how the Microduck learned to walk.
What “Sim to Real” Means?
A simulator is a virtual world with gravity, friction and all the basic rules of physics. Inside it, the robot has a digital twin, which is a virtual copy of itself.
The twin practices walking over and over. It gets rewarded when it moves well and penalized when it falls. This method is called reinforcement learning.
Once the twin gets good, the learned skill, called a policy, is copied onto the real robot. That jump from the virtual world to the physical one is called “sim to real.”
On its official Microduck page, Pollen describes the loop in four steps. Train in simulation, deploy on the robot, refine the simulation, then publish the policy for others to use.
Why Practicing in Simulation Makes Sense?
Here is how the two worlds compare.
| In the Simulation | In the Real World | |
|---|---|---|
| Number of robots | Thousands of virtual copies at once (4,096 in Pollen’s training setup) | One robot |
| Cost of a fall | Free | Possible damage and lost time |
| Practice time | About 1 to 2 hours for a usable walking gait on one GPU, as Pollen reports | Slow trial and error |
Those numbers come from Pollen’s public training code, which uses a learning method called PPO and exports each trained behavior so the real robot can run it.
The Tools Behind the Practice
Two toolkits show up in this story, so let me keep it simple.
Pollen’s public training code is built on MuJoCo, a popular physics simulator. NVIDIA says Hugging Face and Pollen are using Isaac Lab, Isaac Sim and the Newton physics engine to build Microduck. In plain words, the duck’s practice gym can be built with more than one tool.
On stage, Huang gave his own version. He said the robots were trained in the cloud on large NVIDIA systems, then refined inside a digital twin running on an RTX Spark computer.
Related read: Training robots on big cloud machines takes a lot of power. We looked at how AI data centers and your electricity bill are connected, and the story might surprise you.
The Random Tricks That Make It Work
Here is a problem. A simulator is neat and clean, but the real world is messy. Floors are slippery. Motors react a little late. Gears have tiny gaps.
So the training code randomizes things on purpose, such as friction and small gear gaps called backlash. This is called domain randomization. The robot practices in a slightly different virtual world every time, so it does not get confused by real-world surprises.
What This Means for the Stage Stumbles?
Remember the robots that tipped over when told to take a nap? That is the sim to real gap in action.
This gap is a well known challenge in robotics, and the Microduck does not magically remove it. What it does is make the problem cheap to study. Anyone can try, fail and learn without a lab budget.
And that brings us to the question everybody is asking. How much does it cost, and can you still buy one?
Fact 5: The Microduck Costs $399, Sold Out Fast and Ships Before Christmas
Can you buy a Microduck today? Probably not. But there is a lot to know before the next batch, so let’s go through it.
The Price
The robot launched at $399 before taxes and shipping. Pollen calls it an introductory price, so it could change later.
For $399, you get everything you need to start playing. Pollen also sells add-on packs.
| Item | Price | What Is Inside |
|---|---|---|
| Microduck robot | $399 | Robot, battery, USB-C cable, game controller |
| Charger pack | $39 | Dual charger and 2 batteries |
| Dev pack | $119 | 3 spare motors, 5 motor cables, 2 batteries, dual charger, 10 NFC tags, screwdriver, screw pack, Hugging Face credit |
| Accessory pack | $39 | Laser pointer, NFC polaroid, 2 rollers, ball, 10 NFC tags |
Those rollers in the accessory pack are what turn your duck into a roller skater. Nice touch.
The Sellout
Demand was strong. As of October 22, the first pre-order window has closed and the robots are sold out. The Rundown also reported that Microduck brought in about $2.6 million in orders on day one, which means thousands of robots.
Stock can change, so check Pollen Robotics’ official store for restocks.
Shipping and Colors
First shipments are expected before Christmas. Orders were open in North America and Europe at launch.
There are four colors: Cream, Graphite, Lavender and Sky.
Things to Know Before You Buy
A good buying guide should also mention the limits. Here are the main specs and one catch.
| Detail | What to Expect |
|---|---|
| Battery | 2,600 mAh, about 1 hour of continuous runtime |
| Storage and memory | 32 GB storage, 1 GB RAM |
| Connectivity | Wi-Fi and Bluetooth |
| Sound | Built-in microphone and speaker |
One hour per battery is short. That is why the $39 charger pack with two extra batteries looks like a smart add-on for anyone planning long play or training sessions.
Small size also has an upside. A large humanoid that takes a tumble raises safety and cost worries. A 25 cm duck can pick itself back up without a big crash.
Fact 6: It Is Open Source, So You Can Teach It New Tricks (and Test It Before You Own One)
This is where the Microduck stops being a toy and becomes a learning platform.
Related read: Open source is a big theme in AI right now, and not only in robots. If you want the model side of that story, read our post on Mistral Large 4 and open weights.
What Is Actually Open?
The robot’s software, the simulation and the full training stack are published for anyone to read, copy and retrain. You can find them in the main Microduck repository on GitHub, under the Apache-2.0 license. The robot software is written in Rust, and the training code is written in Python.
People noticed. The repository passed 9,000 stars on GitHub within weeks of launch.
One honest note. Pollen has asked people not to call Microduck open source hardware, because the mechanical and electronic design files are not open in the same way as the software. So the accurate phrase is open source software on an affordable robot.
Try It Before You Own One
You do not need the robot to start. Pollen hosts a Microduck simulator on Hugging Face, which is the virtual twin the ducks were trained on. You can drive it in your browser.
That is a great way to see if this hobby is for you before spending $399.
What It Can Do Out of the Box?
The Microduck comes with seven trained moves, including these.
- Walking with a smooth gait
- Sitting down and standing back up
- Kicking a ball
- Grabbing small objects with its beak
- Roller skating (with the skates attached)
- Getting back up after a fall, even from flat on its back
You can control it with the included game controller or have it follow a laser dot. The software also supports Wi-Fi, Bluetooth and live camera streaming.
A Duck With No Words
Here is a fun detail. Unlike Pollen’s Reachy Mini, the Microduck does not speak in words. It communicates through sounds, and each robot gets its own “audio identity” that stays permanent after setup.
Pollen also says using several robots together is about ten times more fun. It even got them playing soccer.
Microduck vs Reachy Mini
Pollen already has a popular desktop robot, so how do they differ?
| Reachy Mini | Microduck | |
|---|---|---|
| Main focus | Human and robot communication | Movement and mobility |
| Movement | Stationary on a desk | Walks, kicks, skates |
| Voice | Speaks in words | Sounds only |
| Best for | A talking desk companion | Training physical behaviors |
Related read: Teaching a robot new tricks is a form of AI training, and AI training is becoming a career path. We covered the details in our post on AI training jobs for experts.
Who Should Get One?
The Microduck is a great fit for students, hobbyists and developers who want hands-on practice with reinforcement learning without a lab budget.
It is not the best pick if you want a talking assistant or a robot that does chores. For that, the Reachy Mini or a future product may suit you better.
Fact 7: This Tiny Duck Is a Preview of Where AI Is Heading, Out of the Screen and Into the Real World
Let’s zoom out for a moment.
For years, AI lived behind glass. It wrote emails, drew pictures and answered questions on a screen. The Microduck is a small sign that AI is stepping off the screen and onto the floor. People call this idea physical AI, which simply means AI that moves and acts in the real world.
NVIDIA’s CEO sees a big shift coming. In the company’s GTC Berlin announcement, Huang said “the ChatGPT moment for robotics is fast approaching.”
Here is why the Microduck fits into that story.
Reason 1: It Makes Robot Learning Cheap
At the same event, NVIDIA showed a research humanoid called Agile ONE SN, scheduled for early 2027. Put it next to the duck and the gap is huge.
| Microduck | Agile ONE SN | |
|---|---|---|
| Height | 25 cm (10 inches) | 1.40 meters (4.59 feet) |
| Weight | Under 2 pounds | 40 kilograms (88 pounds) |
| Movement hardware | 15 motors | 31 degrees of freedom |
| Built for | Learning, play, hobby projects | Frontier humanoid research |
| Availability | Sold out, shipping before Christmas | Early 2027 |
The humanoid is for research labs. The duck is for anyone curious. Both teach the same big lesson, which is that robots can now learn skills in simulation and carry them into the real world.
Reason 2: The Tools Are Reaching Everyone
The NVIDIA Isaac platform got a major expansion at GTC Berlin, including Isaac Lab 3.0, which is now generally available for training robot skills.
NVIDIA is also pushing hardware downward. In August, it announced Jetson Orin Nano 2, an entry-level robotics computer with twice the inference performance of its predecessor and 40 percent less power use at the same performance. NVIDIA says more than 3 million developers have built on its robotics stack.
Put the pieces together. Cheap robots, open training code and affordable chips. That is how a new field grows, because students and hobbyists can finally join in.
Reason 3: The Big Race Is Already On
The Microduck is not the only robot in the headlines. TechRepublic reports that Tesla raised Optimus production from a few dozen robots a week in the second quarter of 2026 to several hundred in August, based on reporting from The Information. Elon Musk wants the robot on sale to the public by the end of 2027, though Tesla has not confirmed a date.
Related read: When robots move from stage demos to factory floors, job worries follow. See what leaders have said in our post on big tech CEOs and their AI job warnings.
The Reality Check
Let’s keep our feet on the ground. The same TechRepublic report notes that Optimus hands have been hard to manufacture, and that the robot has struggled with unfamiliar tasks without a lot of training.
Even our little duck showed the limits live. Three robots fell over on a simple “nap” command, and the battery lasts about an hour.
Here is a fair way to see it.
| Real Today | Still Ahead |
|---|---|
| A 25 cm robot that walks, gets up and skates | Robots that handle unfamiliar tasks reliably all day |
| Robots trained in simulation, then deployed | Humanoids ready for everyday public sale |
| Open training code anyone can use | Tesla Optimus for the public (a target, not a confirmed date) |
So the honest take is this. Exciting, early and moving fast.
Related read: Curious where experts think all of this is heading? Our post on AI 2027 predictions walks through the forecasts.
Frequently Asked Questions
What is the NVIDIA Microduck robot?
It is a 25 cm (10 inch) walking robot with 15 motors, a camera, a depth sensor and a beak that picks up objects. It was built by Pollen Robotics, a Hugging Face company, and shown on stage at NVIDIA’s GTC Berlin 2026.
Did NVIDIA build the Microduck?
No. Pollen Robotics and Hugging Face built it. NVIDIA says they used its Isaac Lab, Isaac Sim and Newton tools in the work, and NVIDIA has agreed to acquire Hugging Face.
Did Jensen Huang present the Microduck at GTC Berlin?
Yes. Huang gave commands to four Microducks on stage, asking them to sit, stand, take a nap and somersault.
How much does the Microduck cost?
The introductory price was $399 before taxes and shipping. Add-on packs cost $39 to $119.
Is the Microduck still available?
As of October 22, 2026, the first pre-order window has closed and the robots are sold out. Check Pollen Robotics’ official store for restocks.
When will the Microduck ship?
The first shipments are expected before Christmas in North America and Europe.
Is the Microduck open source?
The software is. The SDK, simulation and training code are on GitHub under the Apache-2.0 license. Pollen asks people not to call it open source hardware.
Can I try the Microduck without buying one?
Yes. Pollen offers a free simulator on Hugging Face that runs in your browser.
Final Thoughts
A robot shorter than a ruler walked onto one of the biggest stages in tech and made the whole room laugh. But behind the falls and the somersaults sits a serious idea. Robots can now learn in simulation, cheaply and openly, and a $399 duck can be the first step for thousands of curious people.
Here are the 7 facts again, at a glance.
| # | Fact | The Key Takeaway |
|---|---|---|
| 1 | Live orders from Jensen Huang | Four robots, funny falls, great recoveries |
| 2 | Tiny but packed | 25 cm, under 2 pounds, 15 motors, a working beak |
| 3 | Not built by NVIDIA | Made by Pollen Robotics, with NVIDIA’s $12.9B Hugging Face deal pending |
| 4 | Learned in simulation | Sim to real training with domain randomization |
| 5 | Price and availability | $399, sold out, ships before Christmas |
| 6 | Open software | GitHub code, free browser simulator, seven trained moves |
| 7 | Bigger picture | A cheap on-ramp to physical AI, with real limits |
What do you think? Would you bring a Microduck home, or does a walking duck robot feel like the start of something bigger? Tell me in the comments, and share this post with a friend who loves tech.
Information is accurate as of October 23, 2026. Prices, stock and shipping dates can change.
More from oversmartai.com:
Worried about what robots and AI mean for careers? Read about the high paying jobs AI will erase first.
Starting out in your career? See how AI is changing entry level jobs for Gen Z.
Following the business side? Check the latest on AI layoffs at Meta, Amazon and Block.
Wondering if the AI money machine can last? Read the IMF warning about the AI bubble.