Nvidia sold 20,000 robot ducks in a day. The real story is how they learn.
A robot duck napped on stage and sold 20,000 units in a day, proving sim-to-real training is ready for the consumer market.
Business · Source: TechCrunch
What happened
Nvidia CEO Jensen Huang brought a flock of Microduck robots on stage at the GTC Berlin conference. The tiny machines waddled around, somersaulted, and played dead for a nap on voice command. They even made little R2-D2 noises during the demonstration. Hugging Face executive Clem Delangue reported that preorders crossed 20,000 units within a single day following the keynote.
These open-source robots are built by Pollen Robotics. Hugging Face acquired Pollen Robotics in 2025 to expand its hardware ambitions. Nvidia then bought Hugging Face in September of this year. This chain of acquisitions makes the Microduck an entirely Nvidia-owned project. It serves as Hugging Face's second hardware release after the Reachy Mini system.
The ducks are scheduled to ship in time for Christmas. Promotional videos show both adults and kids playing with the devices. But the real engineering magic happens under the hood. The robots learned basic mobility skills inside virtual simulations using the Nvidia RTX Spark platform before ever operating in the physical world.
Key facts
- 20,000 — Preorders placed within a single day of the keynote
- 2025 — Year Hugging Face acquired Pollen Robotics
- RTX Spark — Nvidia platform used to train the robots in simulation
Why it matters
Hardware is no longer the primary bottleneck for robotics development. You do not need to build a physical prototype to teach a robot how to walk or sit. You build the brain in a digital simulation using infrastructure like the RTX Spark platform. You train the model in the virtual world until it masters the behavior. Then you simply download that digital brain into a physical body. This workflow drops the cost of robotics research to near zero for software engineers.
Open-source robotics will create a massive user-generated market for physical behaviors. Owners can train their Microducks to do new tricks and distribute them across a shared user community. We are moving from downloading apps for our phones to downloading physical actions for our robots. The companies that own the simulation platforms and the distribution networks will control the next decade of consumer hardware.
For builders
Sim-to-real is the new standard
Stop testing robotic movements on physical hardware. Builders who use virtual environments to train models will move faster and break fewer expensive parts. Nvidia is positioning its stack as the default tollbooth for this entire workflow.
Open-source hardware drives adoption
Pollen Robotics made the Microduck architecture open source. This allows users to program custom tricks and share them globally. Founders building consumer hardware should leverage community-driven development to keep users engaged long after the initial purchase.
AI needs a friendly face
The AI industry just faced massive backlash over a string of cyberattacks. A cute robot duck acts as the perfect palate cleanser for a skeptical public. Consumer AI products that feel approachable and harmless will win public trust faster than faceless software.
My take
I love this because it proves the future of robotics is just software. The duck did not learn to somersault in the real world. It learned in a simulation, and Nvidia just proved they own the entire pipeline from virtual training to physical deployment.
Original reporting: TechCrunch. This is my rewrite and opinion.