Robotics researchers have succeeded in inducing self-organising flocking behaviour in drones.
Although demonstrated many times using computer models, the latest research – led by Gábor Vásárhelyi of the Hungarian Academy of Science – marks perhaps the first time drone-flocking has been achieved in the real world without the use of a central control system.
The achievement points the way to forward to using drone flocks in a range of applications, from search-and-rescue to mapping and defence.
Flocking behaviour is, of course, commonplace in the natural world, particularly among birds, fish and insects. Understanding its dynamics, however, continues to present a challenge, with different models being suggested.
Nevertheless, understood or not, it happens, often spectacularly. Large numbers of independent organisms are able to move collectively, maintaining direction and not smacking into obstacles or each other.
The same, until now, cannot be said for drones – or, at least, multiple drones that do not respond to a single controller.
One of the main reasons for this, explain Vásárhelyi and his colleagues, is that drone-flocking is primarily the pursuit of computer modellers. Theoretical frameworks for the design of distributed flocking algorithms are all well and good, but they fail to account for real-world conditions.
These, the researchers write, include “constrained motion and communication capabilities, delays, perturbations, or the presence of barriers”.
The absence of such factors from theoretical models limits their value. Barriers and obstacles are not merely isolated challenges, but things that have large and continuing effects on the collective behaviour and cooperation of the flock. Because of this, the scientists note, models that work gracefully on computer screens “tend to oscillate and destabilise quickly under real-life conditions when delays, uncertainties, and kinematic constraints are present”.
The researchers reference a number of apparent real-world examples of autonomous drone flocking behaviour, including events staged by the US military and the band Metallica, but suggest that each is in some way more apparent than real. The drones are all separately programmed to follow specific flight paths, for instance, or instructed to flock towards a specific target, thus limiting variables.
The general principles that govern flocking behaviour, whether in birds or drones, are well understood and uncontroversial. They arise from the interaction of thee simple rules: the need to not crash into a neighbour, the need to steer in the same direction as a neighbour, and the need to follow the average position of a neighbour.
So far, so elegant, but while each of a thousand flying starlings, for instance, is perfectly capable of flying around a tree, a dozen autonomous drones confronted by the same obstacle are likely to crash into it, collide with each other, or fly off in several different directions.
“Creating a large decentralised outdoor drone swarm with synchronised flocking behaviour using autonomous collision and object avoidance in a bounded area is as yet an unresolved task,” note Vásárhelyi and his team.
To a significant extent, however, the new work solves many of the issues.
To do so, the researchers began with modelling that included several extra variables intended to reflect unpredictable real-world problems. These included not only the presence of obstacles and boundaries encountered while moving at high velocities, but also the sudden failure of sensors and short-range communication equipment.
The model was subjected to a process known as evolutionary optimisation – running the program through many generations so that optimally fit features could be identified.
The proof, however, could only be in the real-world robo-pudding. To test their findings the researchers used 30 quadcopter drones and set them up in a physical environment full of obstacles. They programmed them to fly autonomously, set up some of the electronics to fail, and let them go.
Through several runs, the drone swarm flocked and flew without problems – especially, as it turned out, at high speeds.
“The model works in a noisy environment, with inaccurate sensors and short-range communication devices, and in the presence of substantial communication delay and with possible local communication outages,” the scientists report.
Vásárhelyi and colleagues conclude that their results need to be subjected to more sophisticated analysis, but already hold promise for improving drone performance in a variety of contexts.
The research is published in the journal Science Robotics.