Jensen Huang Tells Robot Ducks to Focus: What the Microduck Demo Really Showed

After a keynote full of processors and data centers, four little waddling robot ducks became the stars. Three took a nap on cue; the fourth kept walking. Huang's efforts to call it back briefly turned a technical demonstration into a comedy act.

Four Microducks developed by Hugging Face and France's Pollen Robotics appeared near the end of NVIDIA's GTC Berlin keynote on October 21. Huang asked them to sit, stand and nap. Reporting describes three lying down immediately while the fourth wandered for a few steps before joining them. Later tumbling attempts were uneven, adding to the onstage humor.

The clip traveled widely, resurfacing in a Korean community on October 24 under a playful headline about obedient ducks. Yet the real demonstration concerned deploying motion policies trained in a virtual environment onto real robots. Their mixed success made the simulation-to-reality gap easy to see.

A Korean discussion post at 9:32 a.m. on October 24 linked to an X video of the interaction. Bridge Economy's October 22 video report also documented the nap and somersault requests. Neither proves that the robots understand language like humans or licenses invented audience dialogue.

Training in simulation lets developers test many candidate movements without repeatedly crashing real hardware. Yet friction, sensor noise and physical contact can still make real-world behavior differ. The ducks' little missteps illustrated the remaining challenge better than a perfectly polished animation could.

Microduck is a roughly 25-centimeter bipedal robot developed by Hugging Face and Pollen Robotics. NVIDIA's October 21 release names Isaac Lab, Isaac Sim and Newton among the development tools. The demo described training in the cloud, evaluation on RTX Spark and deployment to hardware. It did not prove robust autonomy in every setting.

Sales and preorder counts vary by reporting date and company statements, so no definitive live total is asserted here. Public developer tools and independently reproduced results will be useful for evaluating how transferable the demonstration really is.