Hugging Face has unveiled a small bipedal robot designed to make training physical AI more accessible. Named Microduck, the robot is shaped like a duck and costs $99; it was developed in collaboration with the French company Pollen Robotics. Despite its toy-like appearance, its primary purpose is to serve as an open platform for developers to experiment with embodied AI and reinforcement learning.

Microduck stands approximately 25 cm tall, weighs less than 800 g, and utilizes 15 motors distributed across its legs, head, and body. The system also includes two inertial measurement units for balance, a camera, a microphone, a speaker, Wi-Fi, Bluetooth, and a small sensor. Its beak is articulated and functions as a sort of gripper, allowing it to manipulate objects.

In demonstrations, the small robot is shown walking, crouching, sitting, following a laser pointer, picking up objects, and even using small roller skates; however, one of its most important features becomes apparent when something goes wrong, as Microduct was designed to fall. Because the machine is small and lightweight, developers can use reinforcement learning to allow the robot to repeatedly attempt a task, make mistakes, and adjust its behavior during training.

When it loses its balance, it manages to get back up and continue the operation. This approach is particularly useful because reinforcement learning typically relies on a vast number of attempts; training a large, expensive humanoid in this manner could pose risks to the equipment itself every time it falls, whereas with a relatively inexpensive platform weighing less than 1 kg, such errors become far less problematic.

The goal is not to compete with large industrial humanoids, but rather to offer a machine small enough to fall without serious consequences and open enough for researchers and developers to modify what it learns; in this context, making mistakes is not necessarily a failure on the robot's part, but rather part of the process used to train it.