A new $399 robot duck is teaching AI to move in the real world

The Microduck is a ten-inch-tall robot equipped with a camera, sensors to map the room, and the ability to balance on its own. It is sold by Hugging Face, a company better known as an online hub where developers share the building blocks of AI. Because the robot is open-source, the company is releasing all its internal software for free. This means that instead of just following pre-programmed commands, owners can use tools to train the duck to learn new physical behaviors, like picking up a pen or rolling around a room.

To teach a robot a physical task, engineers use a process called reinforcement learning. Think of this like training a puppy. Instead of giving the robot a rigid list of instructions for every possible movement, you give it a goal and a reward system. The robot practices thousands of times in a digital simulation—a virtual version of your living room—where it gets a positive score when it moves correctly and a penalty when it falls over. Because it can process thousands of these attempts in the digital space very quickly, it learns to associate certain actions with success. Once it has mastered the move in the digital simulation, that set of instructions is copied onto the physical robot. It then uses its sensors to interpret the real world and apply the movements it practiced virtually.

For most of us, AI is currently something we type into or talk to. The Microduck represents a shift toward physical AI, where the software is tasked with interacting with gravity and solid objects rather than just words. By making the code open, the company is allowing anyone to experiment with how these machines learn. While it is certainly a toy for now, it is also a small-scale laboratory. It forces us to ask what happens when the digital brains behind our chatbots finally get feet, eyes, and the ability to pick things up. Whether this leads to a helpful robot helper or simply a very expensive remote-controlled duck remains to be seen.