The Duck That Learned to Walk in a Thousand Worlds

Anuradha Raghuwanshi — Sat, 31 Oct 2026 22:43:12 +0000

By Anuradha Raghuwanshi

It was past midnight at the Hacktoberfest hackathon, that odd hour when the laptops run hot, the snacks run out, and every conversation drifts somewhere unexpected.

Ours drifted to NVIDIA. Someone mentioned the robots it had just launched, and pointed out that they had been trained across thousands of simulations but hadn’t really been run in the real world yet. Then someone else asked the question that stuck with me: what happens when all that practice finally meets an actual floor?

Nobody had a good answer. Later that night the others went back to their code, but I couldn’t let it go. I opened a new tab and fell down the rabbit hole. This is what I found, and it starts on a stage in Berlin.

Four Tiny Robots, One Big Stage
At GTC Berlin on October 21, 2026, four small robots stood in a row. They were Microducks, small enough to hold in one hand. NVIDIA’s Jensen Huang asked them to do somersaults. Then he told them to take a nap. Three keeled over together, and the fourth needed some persuading. The crowd laughed.

It was a cute moment, but the more I read, the more I felt the real story wasn’t on that stage. It was everything that happened before it.

Meet the Duck
Microduck stands 25 cm tall and weighs under 800 grams. It comes from Pollen Robotics, the robotics team at Hugging Face. Pre-orders opened on August 27 at an introductory price of $399, and the first ducks are meant to reach homes before Christmas 2026 in North America and Europe.

Out of the box it knows seven moves: walking, sitting and standing, kicking, grabbing things with its beak, roller skating, and getting back up after a fall. There’s even a game controller in the box, so you can play with it before writing a single line of code.

And it’s open source. The software uses the Apache 2.0 license. The hardware designs are non-commercial, so nobody can copy the duck and sell it.

A Thousand Ducks, One Afternoon
Here’s the part that made me sit up. Nobody teaches Microduck to walk by letting it crash onto a real floor again and again. That would take forever and wear out the motors.

Instead, the duck learns in a simulator. The project’s training guide shows how to run 4,096 virtual ducks at the same time. Each one stumbles, falls, and tries again, all in parallel, and after one to two hours the group has worked out a walking style that actually works.

NVIDIA’s version of the story has a neat chain to it. An AI agent builds the obstacle course. The virtual duck trains in Isaac Sim, heavier jobs run in the cloud, and the finished skill is checked on an RTX Spark laptop before it goes to the real robot. In the published setup, the duck reads 61 numbers about itself and its surroundings and uses them to control 14 joints.

Picture that for a second: thousands of ducks learning in one afternoon, in a world that doesn’t exist.

Then Reality Walks In
The trouble starts when the duck leaves the simulator.
A robot can look perfectly balanced on a screen and still fall flat on a real floor. Real motors react a little late. Batteries drain. Floors can be slippery or sticky. Parts wear out. A simulator is tidy, and a living room is not.

The team behind Microduck clearly knows this. Their training recipe models how the motors really behave, randomizes the conditions so the duck never gets used to one perfect world, and even simulates the tiny bit of play in each servo’s gears. NVIDIA is also honest that differences between simulation and real life need more refinement before the robot is ready for the real world.

The Part That Hooked Me
This is where I found my answer to that midnight question, or at least a good piece of it.

Training and real life are two different places. The learning happens on a powerful graphics card, with thousands of ducks running at once. But the finished skill is exported and handed to something much smaller. The duck’s brain is a modest Rockchip chip, essentially a quad-core Arm processor with a small graphics unit, and it works completely offline. Whatever the duck learned in that giant virtual crowd has to squeeze into that tiny brain and still work.

The clean version of reality isn’t the real one. The simulated duck lives in a tidy world, so the team deliberately adds mess to it: realistic motor physics, random changes, and gear play. It’s like practicing a speech with a noisy crowd so you’re not shaken on the day.

Nobody has shown the final score yet. Observers of the Berlin demo noted that the jump from simulation to the real robot takes only a few seconds in the video, and that the demo doesn’t measure how well it carries over. Another review points out that we still don’t know how often Microduck succeeds across repeated real-world trials.

So the duck has been trained. It just hasn’t been truly tested yet.

Where the Real Training Begins
When the first ducks reach homes this winter, a second kind of training starts. Each owner’s floors, rugs, and habits will show what the simulator missed. Owners can then adjust the simulation, retrain, and send the improved skill back to the robot. The tricks they teach can be shared with the wider community, so one person’s fix becomes everyone’s upgrade.

That’s what I find most beautiful about it. The duck leaves the factory with seven skills, but its real education happens afterwards, in the open, with thousands of people helping each other.

What the Duck Taught Me

When I finally closed my laptop that night, I had my answer for the room. Practice makes a robot ready, but only the real world shows whether it truly is.

On that stage in Berlin, the fourth duck eventually took its nap. It needed a little persuasion, and that’s exactly the point. In simulation, a thousand ducks can learn to walk in an afternoon. In life, one duck on one floor has to prove it.

Sources
GTC Berlin Microduck coverage: https://pasqualepillitteri.it/en/news/26263/huang-unveils-microduck-robots-gtc-berlin-rtx-spark
Pre-orders, pricing and the stage moment: https://t2online.in/tech/tech-news/pollen-robotics--microduck-can-waddle--roller-skate-and-get-back-up-when-it-falls--and-owners-can-teach-it-more/2008824
Specs, licensing and delivery plans (The New Stack): https://thenewstack.io/hugging-face-microduck-robot/
Training across 4,096 virtual ducks (The New Stack): https://thenewstack.io/?p=22834765
NVIDIA walkthrough on training Microduck in simulation: https://developer.nvidia.com/blog/?p=124028
Sim-to-real limits (TechRepublic): https://www.techrepublic.com/article/news-nvidia-microduck-robot-ai-walking/
Microduck sim-to-real training recipe: https://gitee.com/ncnynl/microduck_rl
Microduck hardware and offline operation (CircuitDigest): https://circuitdigest.com/news/this-microduck-robot-based-on-an-rk3566-has-no-arms-but-picks-things-up-using-ai-and-reinforcement-learning

DEV Community: Anuradha Raghuwanshi

The Duck That Learned to Walk in a Thousand Worlds