Artificial hummingbird flies with the help of machine learning

Scientists have developed an artificial hummingbird weighing 12 grams that can mimic the extreme maneuvers of real hummingbirds thanks to machine learning-based training.

The ‘hummingbird robot’ has two actuators and can therefore move both wings separately from each other. According to the scientists, the fluttering happens at a frequency close to that of real hummingbirds: 40Hz. The artificial bird itself weighs twelve grams but can lift a weight of 27 grams. The model that the researchers at the University of Purdue have developed still draws power via a cable, but in time the model should be able to contain a battery thanks to its carrying capacity.

The bird cannot yet ‘see’, but the electrical currents generated when hitting environmental objects can be used to detect surfaces. The researchers chose the hummingbird because of the animal’s special aerodynamic properties and its ability to perform extreme maneuvers, such as being able to turn 180 degrees with a minimal number of wing flaps.

For regular flying the researchers used a model-based nonlinear system, but for the more extreme acrobatics they had the system without a model reinforced based on simulations. The bird learned more on the basis of trial-and-error.

The research team also developed smaller, insect-like robots weighing up to 1 gram that learned to move based on the algorithms. The wings of those models had to move with a higher frequency. The ultimate goal is to arrive at robots that can be used in various circumstances, such as rescue operations.

The research has been published under the title Learning Extreme Hummingbird Maneuvers on Flapping Wing Robots.