Microduck: The Cute Robot Duck From NVIDIA’s Keynote (and What’s Inside It)

OK, this one is just too cute. During NVIDIA’s GTC Berlin keynote this week, Jensen Huang started playing with these little robot ducks called Microduck, and honestly I loved it.

They waddle around, stomp their feet, make R2-D2 noises, sit down, take “naps” and even do somersaults when you ask them 😂 The video went viral, and according to Hugging Face CEO Clem Delangue, they passed 20,000 pre-orders the next day. This might be the toy of the Christmas.

But I wanted to know what’s inside. Is NVIDIA’s CEO playing with a duck full of NVIDIA chips? Not really. Let’s look.

Which CPU and GPU does it use?

Inside, the duck runs on a Rockchip RK3566, a small, cheap, low-power chip you usually find in single-board computers and budget tablets. So no, there’s no NVIDIA chip in the duck: no Jetson, no GeForce.

How does it compare?

To get an idea of how small this chip is, I put it next to a Raspberry Pi 5, NVIDIA’s own robot board (the Jetson Orin Nano) and my main laptop, the MacBook Pro 14" with the M1 Pro:

The Cortex-A55 cores in the duck are the small “efficiency” type you find in phones, while the Raspberry Pi 5’s A76 cores are much faster. And the Jetson has more than 80 times the AI power. Even my old HP Pavilion from my first videos has 12 times more RAM than this duck.

But that’s actually smart: a cheaper, cooler chip means a cheaper toy, a battery that lasts, and a duck that doesn’t burn your hands. It doesn’t need to be powerful because the hard part happens somewhere else…

So where does NVIDIA come in?

In the training. The duck’s moves (walking, sitting, standing and so on) were learned in a simulation using NVIDIA’s technology, and then moved to the real robot. It comes with seven moves ready to use, no coding needed.

If you want to teach it new tricks, Pollen’s open-source training tools use MuJoCo Warp and reinforcement learning (basically the duck tries thousands of times in the simulation and learns from its mistakes), and they need an NVIDIA graphics card with CUDA. According to the project, a walking gait takes about 1 to 2 hours to train. No NVIDIA card? One extra option sends the training to Hugging Face’s cloud instead. Then you can share your new tricks with other owners.

The rest of the specs

• Size: about 25 cm tall, under 800 g
• Motors: 15 across the legs, head and neck, plus a beak that can grab small things (about 100 g)
• Sensors: front camera, an 8×8 depth sensor, two motion sensors (body and head), microphones and a speaker
• Connectivity: Wi-Fi, Bluetooth and NFC
• Battery: removable 2600 mAh camera-style battery, about 1 hour of play
• Software: open source on GitHub. The hardware design isn’t.

Pollen says some specs are still provisional, like the camera resolution and the depth sensor range, so things could change before it ships.

Who’s behind it

Microduck is made by Pollen Robotics, a French company from Bordeaux that Hugging Face bought in 2025. It’s their second robot after Reachy Mini, a little desktop robot.

And there’s a bigger story in the background: The Information and Business Insider have reported that NVIDIA agreed to buy Hugging Face for around $13 billion. Neither company has confirmed it, so for now it’s just a report. But if it happens, NVIDIA would literally own the duck 🦆

Is it worth $399? For a toy with about an hour of battery, it’s not cheap. But as an easy way to play with real robotics and AI, with open-source software and tricks you can train yourself, it’s a lot cheaper than building your own robot.