The average person probably hasn’t given it much thought, but there are actually tons of fantastic applications for a swarm of drones. From practical operations like crop spraying to a lively light show, the sky is certainly the limit. But first, we have to teach them not to crash into each other.
Enrica Soria, a mathematical engineer and robotics PhD student from the Swiss Federal Institute of Technology Lausanne (EPFL), cares about this issue too. She built a computer model that could successfully simulate the trajectories of five autonomous drones flying through a thick forest without a single collision. However, she realized that in order to test this out in the real world, she’d need to overcome a surprising obstacle: trees.
Drones, especially the higher-end quadcopters she wanted to use, are pricey, and sacrificing a few of them during the test wasn’t exactly ideal. So Soria created a fake forest with soft trees, which were actually just some collapsible play tunnels from Ikea. Soria said that “Even if the drones crash into them, they won’t break.”
Beyond stopping the destruction of costly drones (or of innocent trees), however, the experiment has larger implications. As autonomous drone swarms become more and more commonplace in all kinds of industries and across so many applications, more training needs to be had to ensure these drones won’t collide with each other (or with people or private property) when they’re out on the job. A reliable control system, like Soria’s, is a necessary and important step.
Currently, autonomous swarms are controlled reactively. This means they are always running calculations based on distance from other items so they can avoid obstacles or each other; likewise, if the drones get too spread out, they’ll detect that and move in again. That’s all fine and well, but there’s still the issue of how long it takes the drone to perform these adjustment calculations on the fly.
Soria’s new “predictive control” algorithm actively works to avoid these slowdowns with better and more efficient planning. With it, they communicate with each other to interpret motion-capture data in real time to create predictions of where other nearby drones will move and adjust their own positions accordingly.
Once she set up the fake forest and ran the simulation, she quickly learned that the drones did not crash and that she didn’t need to invest in the softer obstacles. Soria notes, “They are able to see ahead in time. They can foresee a future slowdown of their neighbors and reduce the negative effect of this on the flight in real time.”
Because of this, Soria was able to prove that her algorithm allowed the drones to move through obstacles 57% faster than drones using reactive controls instead of the prediction algorithm. She noted the impressive results in an article published in Nature Machine Intelligence in May.
This project, like many others designed to train autonomous vehicles, was inspired by nature. Yep, like schools of fish, flocks of birds, and swarms of bees. And of course (at least right now), nature is much better at it than we are. Soria notes that “biologists say there’s no central computer,” meaning no single animal or insect directs movement for the rest of the group. Rather, each individual computes its own surroundings—like obstacles and even other fish or birds or bees—and moves accordingly.
Though the concept of predictive control is a first for drones, it’s an old idea. Previously, scientists have used the model to navigate areas and systems for two vehicles moving along predefined trajectories. Predictive control relies on multiple real-time calculations, and if the algorithm running it isn’t elegant, it could max out each drone’s computational capacities.
With so many variables like speed and distance in play, the algorithm also needs to be carefully and thoroughly thought out. Basic parameters like the minimum allowed distance between drones need to be included, to avoid drone-on-drone collisions, but more complex things like no-fly zones and efficient pathway mapping at desired speeds need to be able to compute on the fly without jamming everything up.
As these algorithms get more defined and, thusly, more powerful, it will be easier for them to perform a wider variety of tasks that are tough or inefficient for humans to carry out, like coordinated deliveries in large metro areas or aerial search and rescue missions. But as it is, Soria’s algorithm is a huge step forward for dronekind.