How Microduck robots learn to stand up after a fall

The Keynote Ducks

Eli: October twenty-first, the closing keynote of NVIDIA's GTC conference in Berlin. Jensen Huang, the company's chief, has four little duck-shaped robots on stage, and he's giving them orders. 'Come here. Stay focused.' They don't exactly obey. One sits. One tries a somersault. One pauses, another wanders off, a couple tumble over. When they lie down for a nap, Huang laughs and urges them to stand back up, as NDTV Profit told it. Whether they did wasn't stated. But one of them, after tipping over, rights itself.

Mara: And that's the moment behind a listener's question. They heard NVIDIA bought a company that built these tiny Microduck robots, and now they get back up after a fall without anyone programming the move. What's the significance of the acquisition, what are we actually seeing, and what can we expect? It's Sunday, October 25th, and today's angle is the fine-print audit of a viral keynote demonstration.

Eli: One duck getting back up on a Berlin stage is a lovely moment. Whether it means small robots like this are ready for real work, or are still clever toys, is a much bigger claim.

Mara: Start with the duck itself. Microduck is a bipedal robot: two legs, no wheels, no tracks. It stands about twenty-five centimeters tall. It weighs a little under eight hundred grams. And there are fifteen motors packed into that body.

Eli: So it doesn't roll anywhere. Every step, it's catching its own balance.

Mara: Like an animal does, which means falling comes with the territory. And the price is toy-sized: three hundred ninety-nine dollars to pre-order, before tax and shipping. Hugging Face's chief, Clément Delangue, called it a tiny open-source robot people can teach new tricks.

Eli: A toy's price tag, under a keynote spotlight. Whether NVIDIA actually bought the people who make it is the first thing to check.

The Paper Trail and the Silicon

Mara: So I went looking for the sale, starting from that twenty-five-centimeter duck. And the sale isn't there. NVIDIA didn't buy Microduck, and it didn't buy Pollen Robotics, the French startup that makes it. According to Tech Times, citing SEC filings, NVIDIA signed a definitive agreement on September second to acquire Hugging Face instead. That's the big open platform where AI developers share their models. The price is about twelve point nine billion dollars. That's roughly thirty-two million Microducks at pre-order price. And the deal still needs approval, with a close expected in the first half of 2027.

Eli: And Hugging Face had already bought Pollen, back in April 2025.

Mara: Right, so NVIDIA reaches the duck through its parent. And Huang tells the deal as Delangue's idea. He said he was honored that Clem came to him while weighing Hugging Face's next chapter, believing NVIDIA would be a great home for the company, its community and the future of open models.

Eli: The duck has its own roots, too. Microduck grew out of Open Duck Mini, Antoine Pirrone's open-hardware BDX-style droid, where the physical designs were shared for anyone to build. And the processor inside today's duck isn't NVIDIA's. It comes from Rockchip, a Shanghai-listed Chinese semiconductor company. Automate.org noted there's 'currently no indication that NVIDIA intends to change that hardware configuration.'

Mara: So NVIDIA's star robot runs on someone else's chip.

Eli: And on software anyone can read. Pollen calls Microduck open-source, programmable and retrainable. The SDK, the software development kit with tools to program it, the simulator and the training suite are all published on GitHub. Only the hardware designs stay closed.

Mara: Which means the stand-up trick is something buyers can retrain themselves.

Eli: So it got up on NVIDIA's stage without anyone hand-coding the move. Mara, how it learned that is your end of the story.

Learning to Stand in Simulation

Mara: It learned by reinforcement learning. That's trial and error with a scorekeeper. The software tries move after move, earns points when it ends up upright, and keeps whatever scored best. For this demo, NVIDIA built the virtual scene in Isaac Sim and trained the duck in Isaac Lab, its open-source robot-learning platform. Then comes sim-to-real transfer: lifting what the network learned in that simulated world and running it on real motors, across whatever the simulator got wrong.

Eli: So nobody animates the stand-up. It's just the habit that scored best.

Mara: Mostly, and the stand-up has a wrinkle. Two lessons run at once. One rewards it for getting up. The other pulls it toward a separate expert that already knows how to stand. But that pull waits until the duck's spinning slows down. Pull too early, mid-tumble, and the learning fell apart.

Eli: Like the duck that toppled during its nap in Berlin.

Mara: That one. An attendee put it this way: 'It fell over taking a nap. Then it just… stood up. The whole keynote hall lost it.' Jensen Huang, Deepak Pathak and Clément Delangue point to a physical Microduck that gets knocked over, comes back upright and resumes walking with no human stepping in. And they say the policy went onto the real robot with no hardware-specific tuning and no adaptation step. Seven moves ship that way, walking, sitting and standing among them.

Eli: Trained in a simulator, dropped onto a real body untouched. That's the claim.

Mara: And it's a very small body to drop it onto.

Eli: Size changes everything about falling. One robotics buying guide rates fall recovery across the field as mixed. Some small robots, like the Booster K1, get up on their own. That one stands about ninety-five centimeters. It costs twelve thousand dollars, thirty times the duck. Full-size humanoids generally need a person and a hoist. Because when a robot of thirty to seventy kilos falls, it can damage itself, the room and anyone nearby.

Mara: A duck under a kilo can afford a lot of falls.

Eli: Which is why a learned get-up matters at every size. What nobody showed in Berlin is how often it works. Rising once for a cheering hall and staying up when someone shoves it are two different tests.

The Push Test Reality Check

Mara: Taylor Kolasinski ran tests in September. He trained his own Microduck walking policy and saved a frozen snapshot of it, a checkpoint, every two hundred fifty rounds of training, so each stage of its learning could be tested on its own. Then he put each snapshot through a tool called Proofload, in a lateral push sweep: shoving the simulated robot from the side at rising speeds to find where it tips. Checkpoint 2250 had a cliff. Pushed gently, it almost never fell, just twice in six hundred tries. Just past point six meters a second, more than half went down. A little faster again, and every single run fell.

Eli: A cliff, not a slope. Fine, fine, then gone.

Mara: Then he changed when in the stride the push landed, drawing it at random across a twenty-step window. And the result flipped: a later snapshot, checkpoint 5999, came out significantly safer in all eight conditions. His takeaway: 'A single benchmark push can't tell you which shape you have, and which one you want depends on how sure you are about the disturbances you'll actually meet.'

Eli: So one clean push on a stage says very little.

Mara: And then the comparison that stings. Microduck's own training already scrambles its simulated world, the terrain, the robot's body, the shoves, so it can't count on one perfect setting. That's domain randomization. Kolasinski also varied friction across his test cells. And against his pushes, the VelStand-trained version fell more in every one of fourteen cells.

Eli: Not one cell where it held its footing better.

Mara: The gap ran as high as fifty-seven points. At one push speed, its fall rate went from about one run in eight to more than six in ten.

Eli: On a duck, that's a nuisance. On a seventy-kilo humanoid, it's the room.

Mara: And off the screen, the maintainers had already written down the symptom. After a real fall onto its back, the robot convulsed instead of rising.

Eli: So there are two readings. Jensen Huang, Clément Delangue and Deepak Pathak read the self-righting as proof that simulation-trained skills are ready for real-world work. Kolasinski, Robocap Asset Management and Samaritan Research read polished demos as hiding hardware, sensor and battery limits. What Kolasinski's pushes show is narrower: training it to stand can cost it footing under a shove. And NVIDIA is putting nearly thirteen billion dollars behind the ecosystem this duck comes from.

The Platform Strategy

Mara: Which is where the duck stops being the point. NVIDIA describes a whole pipeline for physical AI, meaning AI that senses and acts in the real world instead of living on a screen: computing, training, simulation, then running on actual machines. One of its leaders put the company's role bluntly: 'We don't build the mechanical robots.' And then: 'What NVIDIA does and NVIDIA does well is we build computers.' He meant chips, systems and software reaching all the way up to the application.

Eli: Rockchip's chip, Pollen's robot body.

Mara: And the duck's training sat on NVIDIA's side of that line. Isaac Lab, the framework that taught it to stand, is built to run thousands of simulations in parallel on NVIDIA's GPUs, graphics processing units designed to crunch massive calculations.

Eli: Thousands of virtual ducks falling at once, none of them breaking.

Mara: Now add Hugging Face. Its community is more than eighteen million people, swapping models, datasets and apps. NVIDIA says it expects the deal to give Hugging Face more resources and support the developers building and sharing open models there.

Eli: Eighteen million developers, and NVIDIA doesn't need one of them to buy a robot.

Mara: Just a way in. One industry observer writing for automate.org called Microduck a relatively low-cost entry point for thousands of developers to experiment with physical AI. He compared its possible role to the Raspberry Pi, and said that wasn't a prediction. The larger prize, he wrote, would be bringing more developers into an ecosystem connecting AI models, simulation, computing and physical machines.

Eli: A toy-priced on-ramp to the whole platform.

Mara: Put those side by side and a shape shows through: NVIDIA's strategy centers on supplying the software, models and simulation pipelines that millions of developers use, and leaves building the robots to partners. Eli, the robots themselves still have to reach the people who ordered them.

Sandbox Reality and Open Signals

Eli: That's the next test. Back to that three-hundred-ninety-nine-dollar price. Pre-orders opened on August twenty-seventh. Tech Times reported more than ten thousand units sold within days.

Mara: Nearly two months before anyone saw one in Berlin.

Eli: Then came the keynote. TechCrunch cited Delangue saying pre-orders passed twenty thousand by the day after it. First deliveries were targeted for before Christmas, with North America and Europe among the launch regions. That's a target, not a promise.

Mara: What does a buyer actually get when the box arrives?

Eli: A duck that, by its own website, works best on flat indoor surfaces. 'It practised on a keynote stage, after all.' Fall recovery and skating improve with software updates. And its removable camera battery runs about an hour, depending on use.

Mara: An hour, then the charger. And the get-up can still fail.

Eli: And the keynote has a footnote of its own. Posts on X from October twenty-first and twenty-second claimed four people were manually operating the ducks while Huang asked them to sit, nap and somersault. That would be teleoperation: humans steering a robot remotely, out of sight. One summary of the thread pinned the claim on an engineer said to work below the stage. The site that wrote that summary said it couldn't verify that person's role or read the posts itself. No outlet has confirmed it, and NVIDIA hasn't addressed it.

Mara: Unconfirmed, unanswered, and sitting right over the moment the hall cheered.

Eli: So, the listener's question. NVIDIA didn't buy the duck's maker outright. It agreed to buy Hugging Face, which owns Pollen Robotics, in a platform deal built around developers. The duck itself is a three-hundred-ninety-nine-dollar sandbox, on a chip from Rockchip, whose learned get-up still has its cliffs. Those four ducks wobbling under Huang's spotlight were showing off a training pipeline, with a toy-priced robot at the end of it.

Mara: And there are four signals to watch. Competition authorities still have to rule on the Hugging Face deal, with a close expected in the first half of 2027. Whether NVIDIA, or anyone independent, ever settles if those stage ducks ran on their own. Real shipping notices for a queue of ten to twenty thousand orders. And whether a future Pollen robot trades Rockchip for NVIDIA's own Jetson chips, which so far nobody has signaled.

Eli: For listeners wanting to dig into the background, look into physical AI.

Mara: Follow technology with us for more stories as machines move from software into the real world.

Eli: That's where we'll leave it. I'm Eli.

Mara: And I'm Mara. Thanks for listening. Until next time on Angle.