Tiny drone uses A.I. to learn from nature’s best pilot, the hummingbird

One of nature’s most remarkable creations is the hummingbird, which flaps its wings up to 80 times per second and which can hover in place and fly in any direction. Now scientists have used machine learning algorithms to study the way these birds fly in order to replicate their abilities in drones.

The robot, developed by researchers at Purdue University, has artificial intelligence (A.I.) which learns from hummingbird simulations and applies its findings to the movements of its flexible flapping wings.

This is useful because of limitations on how small a drone can be made. When drones are shrunk to very small sizes, they cannot generate enough lift to move their weight. But hummingbirds move with a high angle of attack, meaning they move steeply through the air, and high lift, allowing them to use an unstable form of aerodynamics to stay aloft.

Copying the flight abilities of hummingbirds is a challenge scientists have been working on for many years. The latest research involved the researchers spending multiple summers observing hummingbirds in their natural habitats and translating their movements to computer algorithms. The robot could then learn from these algorithms and work out the most effective ways to use its wings.

The robot is made from carbon fiber and membranes which are cut with lasers, and its body is 3D-printed. This means the hummingbird robot weighs only 12 grams (less than half an ounce), and the researchers also created an insect-sized robot which weighed just 1 gram.

The small size and quiet operation of the robot makes it well suited for covert operations as well as search and rescue operations. It could even be used to study real hummingbirds up close in their natural environment.

As well as learning how to move efficiently, the robot also senses its environments by touching surfaces. It does not have visual sensors, but it uses changes in electrical current to track its movements around objects.

“The robot can essentially create a map without seeing its surroundings,” Xinyan Deng, an associate professor of mechanical engineering at Purdue University, explained in a statement. “This could be helpful in a situation when the robot might be searching for victims in a dark place — and it means one less sensor to add when we do give the robot the ability to see.”

A paper describing the robot is available to view on pre-publication archive arXiv and open-source simulations of the flapping wings are available on GitHub.