Did you hear the one about the fly on the treadmill?

When you think of a robot, what comes to mind? Is it human-like, with a dead-eyed stare that makes a shiver go down your spine, like the infamous Sophia? Perhaps they are devious, like those in The Matrix who eventually rise to enslave humanity for power? Or maybe it looks something like the Mars Rover, controlled by a human from afar. Popular imaginations tend to shift towards the fantastical when in reality, a lot of progress is being made in trying to replicate much smaller and simpler organisms, which is precisely what DeepFly aims to do.

DeepFly is a motion-capture software created in 2019 that uses deep-learning (a vein of machine learning) to discern patterns for decision-making in flies, focusing on Drosophila melanogaster, the common housefly. But why flies? Good question. Humans don’t generally like flies. They may be an essential part of the food chain, but they are also irritating and associated with uncleanliness. However, if you look at the fly from a design standpoint, they are ingenious. They can fly or crawl almost everywhere and have small, sticky pad-like structures on their feet that give them the ability to land and rest in any position. Flies, for all they are worth, navigate their environment impressively well, and this is of particular interest to robotics engineers who would like to replicate a fly’s mobility.

Professor Pavan Ramdya and Professor Pascal Fua at the Ecole Polytechnique Fédérale de Lausanne led the study that created and used DeepFly. They set up a series of motion capture cameras around and under a fly positioned on top of a small ball suspended in water, essentially placing the fly on a small treadmill to study it. The fly was able to move on top of the ball without restraint, and therefore acted and made decisions about where to walk as if it were operating freely. The DeepFly team modelled the movements of the fly captured by the cameras surrounding the tiny treadmill and learned how, and possibly why, the fly made decisions in navigating their environment. 

DeepFly allows us to reconsider our traditional interpretations of robotics and their potential for engineering and design. As DeepFly tries to replicate and reimagine the intricacies of nature, they connect the study of how organisms move and operate to the evolving science of robotics. We are in the midst of an unprecedented, interdisciplinary movement in which we are trying to learn from and replicate the intricate beauty of nature. DeepFly is an innovative technology at the forefront, connecting biology, ecological modelling, and robotics.


Ecole Polytechnique F̩d̩rale de Lausanne. (2019, Oct 9). РDeep3DFly: The deep-learning way to design fly-like robots. РScience Daily. https://www.sciencedaily.com/releases/2019/10/191009103256.htm

Photo by Courtenay Crane