Nvidia's Robot Ducks Stole Berlin and 20,000 Preordered
By Daniel Rao, TechPulse USA — covering the companies, models, and money moves shaping artificial intelligence. We read the fine print so you don't have to.
The most important product launch at Nvidia's GTC Berlin keynote last week was not a chip. It was not a server rack, a model, or a partnership slide with too many logos on it. It was four ducks, each about as tall as a wine bottle, waddling around on a stage in Berlin while the most powerful man in artificial intelligence tried to get them to behave. One of them spent the finale facing the wrong way. By the next day, 20,000 people had preordered one.
That sentence should embarrass everyone who writes about technology for a living, present company included. We spend our days solemnly tracking parameter counts and interconnect speeds, and then a $399 robot duck that falls over on command outsells the entire discourse. But the Microduck story is not really about a cute toy. It is about what happens when the hardest problem in AI — getting machines to move competently in the physical world — gets packaged at the price of a mid-range phone, and sold by the man whose hardware trains most of the industry. The ducks are the demo. The demo is the strategy.
The nap that stole the keynote
On 21 October 2026, Huang walked out at GTC Berlin with four Microducks and started giving them orders, according to Reuters' account of the keynote. They sat. They stood. They stomped their little feet and made electronic chirps that more than one report compared to R2-D2. Then Huang told them to take a nap, and three of them keeled straight over onto the stage floor. The fourth needed extra persuading — and when Huang wrapped up his speech, one duck was facing the wrong way entirely.
"Pay attention, stay focused," Huang told it, Reuters reported. The clip went everywhere within hours, which is the part nobody at Nvidia will have needed to plan. A flawless robot demo is a press release. A robot demo in which one machine ignores the chief executive of Nvidia is a story people send to their friends.
When Huang asked for somersaults, the record gets even better, and more honest: a couple of the ducks managed one, while others needed more encouragement, according to Interesting Engineering's account of the keynote. One wandered off, tripped, and got itself back up again. That imperfection is the whole point, and it is also the most interesting thing about the footage. The ducks were not performing a script. They were running learned policies, live, in front of an audience, and learned policies sometimes look like a toddler learning to walk. Because that is, in a meaningful sense, what they are.
TechCrunch's Julie Bort, who covered the preorder surge on 23 October, put the consumer reaction more bluntly than any analyst note could: "It's hard to watch this video without ordering one." She was right, and the numbers backed her up. Hugging Face co-founder Clément Delangue said preorders had passed 20,000 units the day after the Berlin demo, TechCrunch reported, and Delangue suggested the duck could end up the best-selling robot of the year. The robots are expected to be delivered in time for Christmas, which would make them, in TechCrunch's phrase, a candidate for the hottest toy of the holiday season.
A note of proportion, before the hype carries us away: 20,000 preorders are reservations, not deliveries, and the figure comes from the company selling the robot, not from an audited sales report. One relay of Delangue's original post on X even quoted it in predictive form — preorders "going to cross" 20,000, probably the next day — which is a reminder to treat the milestone as a company claim, not a certified count. The distinction matters, and we will come back to it.
What exactly is a Microduck?
Microduck is a 25-centimetre bipedal robot built by Pollen Robotics, the Bordeaux-based robotics team that Hugging Face acquired in April 2025. It is Hugging Face's second robot, after the desktop companion Reachy Mini. It weighs around 800 grams, runs on 15 motors spread across its legs, head and neck, and sees the world through a front camera, a compact depth sensor and two motion sensors. Its beak is articulated: it can dip to the ground, scoop up a small object and pop back upright.
Out of the box, it ships with seven trained moves, according to Pollen's own product page: a velocity-tracking walking gait, sitting and standing, kicking, grabbing, roller-skating (when the skate attachments are fitted), getting back up after a fall, and a quacking voice. Swap its feet for wheels and it roller-skates. Point the laser pointer from the accessory pack at the floor and it will chase the dot. None of this requires writing code — the robot ships with a game controller in the box, alongside the robot itself, a battery and a USB-C cable.
The price is the part that rearranges the market. Preorders opened on 27 August at an introductory $399, before taxes and shipping — a figure Pollen's own page still lists, as we checked before writing this. That is a fraction of what trainable humanoid kits from the United States and China cost, which typically start around $1,700 and climb into the thousands, as New Atlas noted when the robot launched. The ducks are not competing with those machines on capability. They are competing on a different question entirely: how cheap can the entry ticket to physical AI get before students, hobbyists and curious adults start buying in?
Delangue framed it exactly that way in his post on X: "It's a tiny $399 open-source robot you can teach new tricks with reinforcement learning," he wrote. "It can walk, pick things up, get back up when it falls, and even roller-skate." There is one qualification worth stating plainly, because Pollen states it itself: only the software is open source. The mechanical and electronic design files are not part of that release. The code is Apache-2.0 licensed and published on GitHub; the duck's body is not a free-for-all.
Trained in a world that doesn't exist
Here is the part of the story that actually matters for the future of robotics, and it has nothing to do with cuteness. Microduck's behaviours were not hand-programmed movement by movement. They were learned in physics simulation — thousands of virtual ducks practising in parallel — and then transferred onto the real robot. Pollen's product page describes the loop in three steps: train in simulation, deploy on the robot in one step, refine, retrain, redeploy, and publish the resulting policy for the community.
The keynote used the ducks to pitch that pipeline running on Nvidia's own stack. According to accounts of the Berlin presentation, the ducks were trained on Nvidia systems in the cloud, refined in a digital twin, and tested on Nvidia's new RTX Spark machines before the policies went onto the real robots. That last detail deserves a hedge: it comes from keynote reporting rather than a technical paper, and Nvidia has every incentive to make its own hardware the hero of its own keynote. But the underlying claim — sim-to-real training, with the RTX Spark positioned as the machine a developer runs the final check on — is consistent across multiple reports of the event, and it is the commercial logic of the whole exercise.
Think about what that means. Nvidia does not manufacture the Microduck, and it does not need to. If a $399 duck convinces a generation of students and hobbyists that robot training happens in simulation on Nvidia hardware, the company has bought itself the cheapest developer funnel in robotics. Every policy a Microduck owner trains and shares is a small advertisement for the stack it was trained on. The duck is not the product, or not only the product. The pipeline is.
That is also why the open-source software matters more than the price. Owners can retrain a walk, teach a new trick, and publish the policy to the community — Pollen's page makes this the centre of the pitch, not a footnote. A robot that only does what it shipped doing is a toy. A robot whose every behaviour is a file you can read, fork and replace is a platform. The difference between those two sentences is the difference between a Christmas fad and a lasting developer community, and it is the question on which Microduck's long-term fate will turn.
Who actually owns the duck?
The ownership chain behind Microduck reads like a corporate nesting doll, and it is worth laying out because the press coverage keeps shortening it in misleading ways. Pollen Robotics, founded in Bordeaux in 2016 as a spinout of the French research institute Inria, was acquired by Hugging Face in April 2025. Nvidia then agreed in September 2026 to buy Hugging Face for about $12.9 billion, according to Dealroom's reporting of the deal. That transaction remains subject to closing conditions and regulatory review — it has not closed — so the precise formulation is that Microduck is made by a Hugging Face team inside a company Nvidia has agreed to acquire, not by Nvidia itself.
Huang used the Berlin stage to make a broader argument about that European connection. "No region has dedicated itself to making mechatronics, electronics and mechanical systems at the level of precision and craft as the European region," he said at GTC Berlin, according to Reuters, adding that "every company in Europe will become an AI company." It was a flattering line for a Berlin audience, and it was also a strategic one: Europe's engineering base, he argued, gives it an edge in physical AI and robotics. The ducks, built in Bordeaux and demoed in Berlin, were the exhibit.
There is a reason to be slightly cynical about the timing, and it is not a secret. The same week as the keynote, the AI industry was absorbing a run of bad headlines about AI involvement in cyberattacks — TechCrunch's own Microduck piece opened by calling the duck video "a palate cleanser" after that news. A viral duck is a very convenient palate cleanser. It does not follow that the demo was staged as a distraction; it follows that Nvidia understands, better than most companies, that public affection is a strategic asset, and that a duck that falls over is worth more of it than a keynote slide about interconnect bandwidth.
Our take: the duck is real, the deadline is the risk
So is Microduck a gimmick? No — and that is a more interesting answer than either the hype or the backlash allows. The gimmick reading fails on the spec sheet: a 25-centimetre, 15-motor biped with a camera, depth sensing and a published reinforcement-learning stack, at $399, is a genuinely capable research platform that happens to look like a bath toy. The hype reading fails on the calendar. Promising 20,000-plus buyers delivery before Christmas means manufacturing, testing and shipping tens of thousands of legged robots — each with 15 motors, a battery and a sensor suite — in roughly two months, from a small French team that has never shipped at this volume. Taking orders is easy. Fulfilling them is the hard part, and it is where most hardware darlings go to die.
The honest verdict is that Microduck is the most important robotics product of the year precisely because it is not trying to be a robot servant. It is not going to fold your laundry — its beak lifts only light objects, and anyone imagining otherwise should read the spec sheet twice. What it might do is put a trainable robot on enough desks, in enough classrooms and bedrooms, that physical AI stops being something people watch in keynotes and starts being something people tinker with. That is how personal computing started, too: not with a machine that did your job, but with a machine cheap enough to play with. Whether Pollen can ship 20,000 ducks by Christmas will decide whether this is a movement or a meme. The ducks, for their part, have already done their job. They got everyone's attention — including, presumably, the one that spent the finale facing the wrong way.
What could go wrong
Three risks deserve a clear-eyed paragraph each, because the preorder headlines skip all of them. First, delivery: first shipments are targeted before Christmas in North America and Europe, but that is a target, not a guaranteed date, and later orders may wait considerably longer. Second, the price: $399 is explicitly an introductory price, before taxes and shipping, so later buyers may pay more, and the accessory packs — $39 for the charger pack, $39 for accessories including the laser pointer and rollers, $119 for the developer pack — add up quickly. Third, privacy: a camera and microphone on a robot that roams your home deserve scrutiny. Open code lets developers inspect what the duck does with what it sees, but most buyers will never read it, and a cute face is not a privacy policy.
None of that makes Microduck a bad buy. It makes it a first-generation product from a small team, sold at a price that invites impulse purchases, in a week when impulse was running high. The smart buyer treats it as what it is: a learning platform with a famous demo, not a finished consumer appliance. The smart observer watches the shipping dates, not the preorder counter.
Quick answers
How much does a Microduck cost?
$399 as an introductory preorder price, before taxes and shipping. Charger, accessory and developer packs cost $39, $39 and $119 respectively, according to Pollen's product page.
When will preordered Microducks arrive?
First deliveries are targeted before Christmas 2026 in North America and Europe. That is a target, not a guaranteed shipping date, and later orders may face longer waits.
Can you really teach it new tricks?
Yes — that is the point. The software stack, including the simulation and reinforcement-learning training code, is published under the Apache-2.0 licence. Owners can retrain behaviours in simulation and share the resulting policies with the community. The hardware design files are not open source.
Is Microduck made by Nvidia?
No. It is made by Pollen Robotics, Hugging Face's robotics team in Bordeaux. Nvidia agreed in September 2026 to acquire Hugging Face for about $12.9 billion, but that deal has not closed. Nvidia's role is the training stack the ducks were developed with — and the keynote stage they went viral on.
Sources: TechCrunch (Julie Bort, 23 October 2026); Pollen Robotics Microduck product page; Reuters' account of the GTC Berlin keynote, as carried by Dealroom and the Economic Times; t2ONLINE (Telegraph) Microduck report; TechRepublic Microduck report; Interesting Engineering's keynote account, as relayed by DeFoor Report.
For more on Nvidia's physical AI push, see Nvidia's plan to ship your AI factory in a shipping container — and for the London side of Huang's European tour, what Huang told London about opaque AI. For a reminder that robots do not always do what their creators intend, the 78-year-old professor versus the burrito bot remains our favourite cautionary tale.
By Daniel Rao, TechPulse USA — covering the companies, models, and money moves shaping artificial intelligence. We read the fine print so you don't have to.
Comments
Post a Comment