NVIDIA CEO Jensen Huang showcases AI-powered Microduck robots at GTC Berlin 2026: How the tiny machines learn to sit, stand and nap on command

Imagine a robot that can walk into a room, follow instructions and decide when it is time to take a nap. Now imagine four of them doing exactly that on stage in front of one of the world's most influential technology executives. That was the rather amusing scene at NVIDIA's GPU Technology Conference (GTC) in Berlin on October 21, 2026, when CEO Jensen Huang introduced an unusual group of robotic assistants. The tiny machines, called Microducks, walked towards Huang before demonstrating their ability to sit, stand and lie down on command. While their movements were far from perfectly synchronised, the demonstration offered a glimpse into how artificial intelligence is moving beyond computer screens and into the physical world. It also brought a lighter moment to a keynote focused on NVIDIA's expanding ambitions in artificial intelligence, robotics and computing infrastructure.

NVIDIA Microduck Robots Follow Jensen Huang's Commands

The four Microducks appeared towards the end of Huang's keynote, walking onto the stage independently before gathering in front of the NVIDIA chief executive. Developed by Hugging Face and French robotics startup Pollen Robotics, the small machines were designed to demonstrate how AI-powered systems can learn and execute physical movements.

Huang began by asking the robots to sit down. He then instructed them to stand up, with the machines attempting to follow his commands. The most entertaining moment came when Huang announced that it was time for a nap.

Three of the Microducks quickly lowered themselves onto the stage and lay down. The fourth, however, continued walking for a few more steps before finally settling beside the others. The unexpected delay prompted laughter from Huang, who encouraged the robots to get back on their feet.

Although the demonstration was simple, it highlighted an important development in robotics. Machines are increasingly being trained to understand instructions and translate them into physical actions, rather than simply responding to commands on a screen.

What Are NVIDIA Microduck Robots?

Microduck is an open-source robotics platform created for developers, students and AI enthusiasts who want to experiment with robotic behaviour. The idea is to make robotics development more accessible, allowing users to teach machines new physical skills, modify their behaviour and share their work with the wider community.

Each Microduck stands approximately 10 inches tall and weighs less than two pounds. Their compact design makes them suitable for experimentation without requiring the space or infrastructure associated with larger industrial robots.

The platform was developed using NVIDIA's Isaac Lab and Isaac Sim tools, alongside the company's Newton physics engine. These technologies allow developers to create virtual environments in which robots can learn how to move, interact with objects and respond to different situations.

Instead of repeatedly testing every movement on a physical machine, developers can first experiment with a digital version of the robot. This approach is particularly useful because training robots in the real world can be expensive, time-consuming and difficult to scale.

How NVIDIA Trained Its Microduck Robots Using AI

Before introducing the robots on stage, Huang showed a video explaining how an AI agent helped train a Microduck to perform physical movements. The process began with the AI agent designing a training course and testing a virtual version of the robot inside NVIDIA's Isaac Sim simulation environment.

The virtual robot could then practise its movements, allowing developers to evaluate its behaviour before transferring the results to the physical machine. The computationally intensive training work was carried out on cloud servers, using NVIDIA's computing infrastructure.

Once the training was complete, the results were checked on an RTX Spark laptop before being transferred to the physical Microduck. The demonstration illustrated how AI agents, simulation software, cloud computing and local hardware can work together to develop robotic capabilities.

It also highlighted the importance of digital twins, which are virtual representations of physical machines that allow developers to test and refine their behaviour.

NVIDIA RTX Spark Laptops And Robotics Computing

Huang also discussed NVIDIA's RTX Spark computing systems, which launched the previous week.

According to the NVIDIA CEO, the systems offer up to one petaflop of computing performance and 128GB of memory.

To explain how much computing hardware has evolved, Huang compared the RTX Spark with NVIDIA's DGX-1 system, which he personally gave to OpenAI in 2016. The DGX-1 was a substantial computing system that required two people to carry it upstairs, Huang recalled.

The RTX Spark, by comparison, is compact enough to be held in one hand. The comparison offered a striking illustration of how computing power has become more accessible over the years, potentially making advanced AI development easier for individuals and smaller teams.

The RTX Spark also played a role in the Microduck training process, with the laptop being used to check the results before they were transferred to the physical robots.

NVIDIA GTC Berlin Highlights Physical AI And Europe's AI Ambitions

The Microduck demonstration concluded a keynote that focused on NVIDIA's plans for Europe, including AI factories, autonomous agents and physical AI. Physical AI refers to artificial intelligence systems that can interact with the real world through machines such as robots, autonomous vehicles and other intelligent devices.

Unlike conventional AI applications that generate text, images or digital information, physical AI must also account for movement, surroundings and real-world conditions. The Microducks offered a small but engaging example of how these technologies can work together.

Their ability to follow basic commands demonstrated the connection between AI training, simulation and physical execution, even when their movements were not perfectly coordinated. Huang ended his presentation with a message encouraging wider participation in AI development.

"Everybody can build agents," he said.

The statement captured the broader idea behind the demonstration. As AI tools, computing systems and open-source robotics platforms become more accessible, the ability to experiment with intelligent machines is no longer limited to large technology companies.

And if four tiny robots learning when to sit, stand and sleep are any indication, the next chapter of artificial intelligence could be just as entertaining as it is transformative.