Optimized flocking of autonomous drones in confined environments

Science Robotics 2018 July Cover

G. Vásárhelyi, Cs. Virágh, G. Somorjai, T. Nepusz, A. E. Eiben, T. Vicsek, Optimized flocking of autonomous drones in confined environments, Science Robotics, Vol. 3, Issue 20, eaat3536 (2018)


We published our first drone flocking results back in 2014. It took yet another four years of thinking, simulations, experiments and hardware/software re-design since then to overcome the major obstacles preventing us from scaling up our drone swarm system both in terms of flock size and flocking speed. In this article we presented a substantially enhanced model of flocking which has also been optimized by an evolutionary algorithm. With this novel solution we were capable of flying 30 drones at 8 m/s speed in confined environments (presented as virtual GPS walls and obstacles) in much smoother self-organization than anyone (including ourselves) ever before.

This work was pretty exciting as it also made us advance in our understanding of the structure and dynamics of complex systems in general. A very general take-home message of this article is that complex systems under real-life conditions must be optimized to hell to function properly and that a complex architecture in a stochastic environment worth nothing without the proper instantiation of its model parameters. Such as a drunken person immediately falls out of the range of “normal behaviour” with slightly modified brain-parameters, a slightly mistuned or not well optimized flock - let it be of birds of a feather or artificial drones - will also fail shortly. Therefore, the eternal beauty of natural (or by now artificial) flocks or other complex systems comes at a high price: the harmony and perfect synchrony of individuals is the result of infinite practice with pain and sweat, with the help of the greatest optimizer of all times, evolution.

Video Abstract

Q & A


We address a fundamental issue of collective motion of aerial robots: how to ensure that large flocks of autonomous drones seamlessly navigate in confined spaces. The numerous existing flocking models are rarely tested on actual hardware because they typically neglect some crucial aspects of multirobot systems. Constrained motion and communication capabilities, delays, perturbations, or the presence of barriers should be modeled and treated explicitly because they have large effects on collective behavior during the cooperation of real agents. Handling these issues properly result in additional model complexity and a natural increase in the number of tunable parameters, which calls for appropriate optimization methods to be coupled tightly to model development. In this paper, we propose such a flocking model for real drones incorporating an evolutionary optimization framework with carefully chosen order parameters and fitness functions. We numerically demonstrated that the induced swarm behavior remained stable under realistic conditions for large flock sizes and notably for large velocities. We showed that coherent and realistic collective motion patterns persisted even around perturbing obstacles. Furthermore, we validated our new model on real hardware, carrying out field experiments with a self-organized swarm of 30 drones. This is the largest of such aerial outdoor systems without central control reported to date exhibiting flocking with collective collision and object avoidance. The results confirmed the adequacy of our approach. Successfully controlling dozens of quadcopters will enable substantially more efficient task management in various contexts involving drones.

Download Full Text

The full article is available online at Science Robotics.

Download Supplementary Material

Download Flight Logs

Flight logs related to the article are accessible on Dryad.

Download Simulation Code

The code basis of the multi-drone simulation that was used in the article is open-source and can be found at github.com/csviragh/robotsim.

Authors and Affiliation

1MTA-ELTE Statistical and Biological Physics Research Group
2ELTE Department of Biological Physics
3CollMot Robotics
4Molde University College
5Vrije Universiteit Amsterdam


Media Coverage