A Minimalist Flocking Algorithm for Swarm

Robots

Christoph Moeslinger1 , Thomas Schmickl1 , and Karl Crailsheim1

Karl-Franzens University Graz, Artificial Life Lab of the Department of Zoology,

Universitätsplatz 2, 8010 Graz, Austria

{christoph.moeslinger, thomas.schmickl}@uni-graz.at

Abstract. In this paper we describe a low-end and easy to implement

flocking algorithm which was developed for very simple swarm robots
and which works without communication, memory or global informa-
tion. By adapting traditional flocking algorithms and eliminating the
need for communication, we created an algorithm with emergent flocking
properties. We analyse its potential of aggregating an initially scattered
robot swarm, which is not a trivial task for robots that only have local
information.

1 Introduction

One of the most amazing developments in biological evolution is the domain

of social insects. These animals, although being very small, achieve impressive
feats. Bees, ants and termites live in elaborately constructed nests which are,
in comparison to the insect, gigantic [1]. The colony super-organism can be
characterized as being swarm-intelligent [2] because its abilities, for example to
optimally allocate foragers to food sources, are a result of the interactions within
the swarm and cannot be achieved by the single individual. This decentralized
and distributed way of achieving a goal is an interesting and useful field of study
which has inspired the fields of swarm-intelligence [3, 4] and swarm robotics [5].
In the last decade, a lot of control strategies and algorithms for robotic swarms
have been presented, both in simulated and real robot swarms.
Innovation and industrial progress make it possible to manufacture such
swarm robots in smaller and smaller sizes [6]. For such small robots, reach-
ing a common goal is not an easy task. This is due to the constraints that come
with decreasing size, like small sensor ranges, very limited computational power,
little memory and imprecise locomotion. Thus, developing control algorithms
for small scale swarm robots raises interesting questions, like ”how can a robot,
which only knows about the very small part of the environment around itself,
achieve a common goal with the rest of the swarm?” and ”given all constraints,
what is the most efficient way to reach a common goal?”.
Reaching such a common goal often implies that the swarm has to be aggre-
gated for cooperative work, transport [7, 8] or assembly [9, 10]. This aggregation
task seems to be relatively trivial, but the constraints of a swarm of small robots
create some conceptual problems. Small robots usually do not have the capa-
bility of long range communication and do not possess global information like
position and heading. When only local information is available, the coordina-
tion of numerous autonomous agents requires a different approach. Examples of
such an approach already exist in nature: the phenomenon of flocks, herds and
schools.
Flocks can be aggregations of up to several thousand individuals which move
together with astounding elegance and flexibility. Craig Reynolds was amongst
the first to abstract this behaviour to steer a swarm of simulated birds which
he called boids [11]. To do this, he implemented three behavioural rules in his
autonomous boids: collision avoidance to evade obstacles and flock mates which
are too close, flock centering to stay close to flock mates and velocity / heading
matching to move in the same direction as nearby flock mates. The resulting sim-
ulated flocks appear very similar to real flocks. As a result, several approaches
to adapt this behaviour to a robotic swarm have been made. Although these
pursues were successful in creating flocks, all of them incorporated something
that is usually not found in real bird flocks or fish schools: communication and
knowledge of the own heading [12–15] or predefined leaders [16]. Other interest-
ing approaches used light beacons (and the resulting shadows) in an arena to
generate a flocking swarm, either using evolved neural networks [17] or adapted
aggregation algorithms in combination with situated communication [18], where
the only significance is the presence or absence of an ”empty” message. Usually,
the matching of velocity and heading was only possible when the robots were
able to communicate with their flock mates through a stable communication
channel. This means that each robot needed to know its own heading (through
a digital compass) and the headings and speeds of all its (near) flock mates.
This involves extensive communication, which can be far too complex for
small and numerous swarm robots, and, more importantly, that is also not in
accordance to how a real flock achieves alignment. In real swarms, the animals
identify the heading of each other because they derive it visually from their
flock mates’ body form, although most fish species also have a specialized lateral
line organ which is additionally used for alignment [19]. Such methods could be
adapted for swarm robots by using on-board cameras and image recognition, but
that would probably be way too complex for very small robots.
We found out that there is also a much easier way to create flocking be-
haviour. By discretizing the robots’ sensor fields into sectors and using different
distance thresholds for attraction and repulsion in these sectors, robot swarms
can achieve emergent alignment.

2 Material & Methods

2.1 Algorithm Requirements

The main aim of our flocking algorithm is simplicity. This means that we devel-
oped our algorithm in consideration of very limited computational power and
only minimalist swarm robot equipment. Such a basic equipment is a set of
distance sensors, which are usually used for collision avoidance. We wanted to
utilize just these sensors to generate complex swarm behaviours.
Swarm robots usually have IR-sensors in all directions which enable the de-
tection of reflecting surfaces like walls, obstacles and other robots. A major dis-
advantage of these IR-sensors is, that they have very limited range and cannot
discriminate robots from obstacles, except when using a combination of active
and passive sensing. Here, active sensing means that the robot activates its IR-
light at the position of the sensor and checks for reflections, whereas passive
sensing means that the robot only checks for IR-light from other robots without
emitting IR-light itself. Usually, the brighter the sensed IR-light, the higher is
the value that the sensor returns. These IR-sensor values are all that is needed
for the effectivity of our algorithm. Of course other distance measuring sensors,
like ultrasonic sensors, can be used for our algorithm as well.
The minimal requirements of the algorithm are 4 circumferential distance sen-
sors with limited range and 3 discrete reactions in movement: move straight, turn
left or turn right. The algorithm does not require any global information about
positions or headings, precise sensor information, memory, elaborate robot-to-
robot recognition or communication.

2.2 Flocking Algorithm

Each robot in the swarm periodically emits IR-pulses. The robots then react
(move straight, turn left or turn right) depending on information from their
active and passive IR-sensors. These sensors are polled periodically and the
returned values are then checked against predefined thresholds (Fig. 1B) in a
simple subsumption architecture (Fig. 1A).
First, the active IR-value for the front sensor is polled to find out whether
there is an obstacle in front. If the value for the reflected IR-light is above a
certain threshold, the robot turns away in a random direction. This is the basic
collision avoidance of our robots.
If there are no objects in its way, the robot checks the passive IR-values of all
sensors. If the front, left or right sensor is above a certain threshold, the robot
turns away from what is presumably another robot which is too close. This rule
is usually referred to as the separation rule in flocking algorithms.
If there is no other robot too close, the robot checks the passive IR-values of
its left, right and rear sensors. For every sensor that returns a value that is above
the environmental IR-light threshold but below the threshold which defines the
maximally desired distance to another robot in that sector, the robot performs a
basic vector addidtion and adds up all turns. It then decides to turn in a direction
depending on whether there were more left or more right turns. Robots in the
rear zone trigger a random turn reaction. This rule is usually referred to as the
cohesion rule in flocking algorithms.
The third rule in flocking algorithms is usually the alignment rule which
generates the common direction of movement in a flock. Since we wanted our
Fig. 1. A: Simple subsumption architecture depicting the flocking algorithm. The first
decision results in collision avoidance, the second decision results in robot separation
and the third decision results in flock cohesion and emergent alignment. B: Simplified
depiction of the perceived IR-values of other objects (reflected active IR) or flock mates
(passive IR) dependent on their distance to the robot. Thresholds and the resulting
zones for a robot with 4 IR-sensors.

algorithm to be as simple as possible we wanted to exclude complex communica-

tion or image recognition procedures and implemented a method which generates
emergent alignment. To achieve this we adjusted the thresholds for the cohesion
rule so that robots tend to follow other robots. This is done by simply shifting
the threshold for the rear sensor more outwards in comparison to the thresh-
olds for the left and right sensors (zones 3 and 4 in Fig. 1B). Depending on the
position and heading of two approaching robots, one robot will be behind the
other robot. When both robots move, the robot behind will turn towards the
robot in front before the robot in front reacts and turns around. This creates a
leader robot and a follower robot, purely by chance. These two robots will then
move around in the arena without separating. If the path of these two robots
is blocked by an obstacle or another robot joins the flock, the arrangement can
change instantly. If two robots approach frontally, they will avoid each other,
only to turn back to each other shortly after, which can create a deadlock situa-
tion. To prevent such situations, we implemented a random-turn reaction which
means that robots will randomly turn either left or right when avoiding other
robots in front (zone 1 in Fig. 1B).

2.3 Simulator

We conducted simulations firstly as a proof-of-concept and secondly to evaluate

the aggregation and flocking capability of a robot swarm which uses our algo-
rithm. These simulations use very abstracted and ideal two-dimensional models
of autonomous mobile agents with circumferential sensors. For our simulations
we used a custom-built simulator (Fig. 2A) based on the multi-agent program-
ming language NetLogo [20]. A short video can be seen at [21].
 The tests were conducted with different numbers of robots in differently sized
arenas. For generality, we assumed 1 robot-diameter as the unit of measurement.
A simulated robot moved with 3 units per second and polled its distance-sensors,
which had a maxmium range of 5 units, 60 times per second. The thresholds
were set to 1 unit for zone 1, 1 unit for zones 2, 2 units for zones 3 and 3.5
units for zone 4. These values were derived from tests with real swarm robots
[22]. To emulate real-world conditions and to avoid certain deadlocks (e.g., two
robots circling each other) we introduced inaccuracies. In the simulations, the
sensor values were subject to a 5% random-normal error in measurement and the
speed of the individual robots was subject to a 5% random-normal variance. We
simulated swarms of 5, 10, 15, 20 and 25 robots in bordered square arenas with
the sizes of 3600, 7200, 10800, 14400 and 18000 robot-diameter2 . The sizes of the
arenas were chosen to create density-neutral setups so that we could measure
the impact of swarm size on the efficiency of the algorithm.

3 Results

Our simulations contribute to two distinct analyses. On the one hand, we wanted
to investigate the flocking capabilities of our swarms and, on the other hand, we
wanted to determine the algorithm’s efficiency to let randomly scattered robots
form an aggregation in an arena. Here, aggregated or flock means that we counted
each robot which was inside the IR-sensor field of another robot and counted
the total number of robots within that connected swarm.

3.1 Flocking Analysis

Our first simulations were set up to find out the mobility of an aggregated robot
swarm. For this we simulated a huge 300x300 robot-diameter2 arena so that
the robots were not confined in their movement by the borders. An already
aggregated swarm of 1 to 25 robots with randomized headings was placed into
the middle of this arena. We then added up the movements of all robots for 60
seconds and calculated the average distance covered by a robot in the swarm.
Additionally, we measured the movement of the coherent flock by calculating its
center of mass and adding up its path.
An ideally mobile flock would cover the same distance as the average distance
covered by a robot in the flock. Since our algorithm does merely impede but not
prevent a flock from splitting up, we only counted simulation runs where the
whole swarm stayed together the whole time of the simulation, which was the
case, on average, in 46% of all runs. Each experiment was repeated until there
were 10 successful runs. A comparison of the distance measurements for the
robots and the center of mass of the flock can be seen in Fig. 2B. The average
movement of a robot in a flock differs slightly due to the implemented speed
error.
 average robot center of mass
200

distance covered [robot-diameters]

150

100

50

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
(A) (B) flock size [robots]

Fig. 2. A: Screenshot of our simulator with a swarm of 10 robots (black cubes) in a

small arena with borders. The path of the center of mass of the flock is depicted by a
grey line. B: Average movement of a robot in a flock (dark grey) in comparison to the
average movement of the center of mass of the flock (light grey) in 60 seconds. Means
and standard deviations of 10 repetitions.

3.2 Aggregation Analyses

General Efficiency: Next, we focussed on finding out the aggregation efficiency

of a robot swarm which uses our algorithm. In these simulations the robots had
been randomly scattered in an arena and had to aggregate. There was no global
information about where to aggregate and the robots reacted purely on short-
range IR-detection. A swarm was considered as being aggregated when at least
60% of its members were in the same flock. We measured the time it took a
scattered swarm to aggregate under different combinations of swarm size and
arena size. 100 repetitions with randomized starting positions and headings were
evaluated, the result can be seen in Fig. 3A.

Density-neutral Efficiency: Further simulations were conducted only under

density-neutral conditions, for example 5 robots in a 3600 robot-diameter2 arena
or 10 robots in a 7200 robot-diameter2 arena. This was done to measure the
influence of swarm size on the efficiency of the algorithm. A supplementary
measurement was made by increasing the desired aggregation size from 60% to
80% (Fig. 3B). Each simulation was evaluated by the median, first and third
quartiles of 100 repetitions.

4 Discussion

The flocking analysis showed that the mobility of a flock of swarm robots which
use our algorithm was dependent on the flock’s size. A small flock of 5 robots still
moved 69% of the distance a single robot could have moved, whereas increasing
the flock size to up to 25 robots decreased the mobility down to 25%. This
decrease is of course a major drawback in comparison to traditional flocking
algorithm implementations, where large flocks can still be quite mobile.
 median, 1st and 3rd quartiles for desired aggregation size
60% 80%
800 1200 2400

time to aggregate (60%) [s]

time to aggregate (80%) [s]

1000
time to aggregate [s]

600 1800
800

400 600 1200

400
200
600
25
200
20 t s]
bo
0 ro
1800
0
15
e[ 0 0
1440
0
0 10 siz 5 robots 10 robots 15 robots 20 robots 25 robots
arena si 1080 rm
ze [rob 7200 5 wa
(A) ot-diam 3600 s (B)
eter 2]

Fig. 3. A: Aggregation speed analysis for all swarm sizes and all arena sizes. We mea-
sured the median time it took for an initially scattered robot swarm to form an aggre-
gation of at least 60% of the whole swarm. B: Comparison of aggregation time for an
initially scattered robot swarm in density neutral setups for 60% (light grey) and 80%
(dark grey) desired aggregation size. Median, first and third quartiles of 100 repetitions.

The aggregation analyses suggested that our flocking algorithm was quite ef-
ficient concerning aggregating scattered robots. It should be noted that, contrary
to our expectations, there was an increase in aggregation time when increasing
the number of robots in the same arena. In retrospective this can be explained
by the flocking analysis, which showed that with increasing numbers of flock
mates, the movement of the flock was more and more limited. Thus it became
harder for two larger flocks to merge in order to attain the desired aggregation
size.
Our density-neutral experiments showed that the aggregation time was in-
creased when increasing the number of robots and the size of the arena. This is
due to the larger distances that flocks have to cover in order to join other flocks
and the aforementioned decrease of flock movement in bigger flocks. When the
desired aggregation size was changed from 60% of the whole swarm to 80% of the
whole swarm, there was a large increase in aggregation time. The reason why we
did not test the goal of achieving 100% aggregation was that, when working with
numerous swarm robots, one has to expect a small fraction of malfunctioning or
stuck robots, so that perfect achievements are rather improbable.
Summing up, our flocking algorithm enables swarm robots to form a coherent
and reactive flock which moves around in the arena randomly. Our results suggest
that it works well with small swarms and is especially suited for robots with
minimal equipment. It could, for instance, be used for the aggregation part of
a more comprehensive scenario. Moreover, the algorithm can also be used for
a heterogeneous robotic swarm when the robot types use the same distance-
sensing method. Our next step will be to port the algorithm to a real swarm of
up to 30 heterogeneous swarm robots [22, 23].
1
Supported by: EU-IST-FET project ‘SYMBRION’, no. 216342; EU-ICT project
‘REPLICATOR’, no. 216240; EU-IST FET project ‘I- Swarm’, no. 507006.
References
1. Wilson, E.O.: The Insect Societies (Harvard Paperbacks). Belknap Press (Novem-
ber 1974)
2. Beni, G., Wang, J.: Swarm intelligence. In: Proc. of the Seventh Annual Meeting
of the Robotics Society of Japan. (1989) 425–428
3. Kennedy, J., Eberhart, R.C.: Swarm Intelligence. Morgan Kaufmann (2001)
4. Bonabeau, E., Dorigo, M., Theraulaz, G.: Swarm Intelligence: From Natural to
Artificial Systems. Oxford Univ. Press (1999)
5. Dudek, G., Jenkin, M., Milios, E., Wilkes, D.: A taxonomy for swarm robots.
Intelligent Robots and Systems ’93 1 (1993) 315–325
6. Seyfried, J., et al.: The I-SWARM project: Intelligent small world autonomous
robots for micro-manipulation. In Şahin, E., Spears, W.M., eds.: Swarm Robotics
- SAB 2004 Int. Workshop, Berlin, Germany, Springer-Verlag (2005) 70–83
7. Trianni, V., Groß, R., Labella, T., Şahin, E., Dorigo, M.: Evolving aggregation
behaviors in a swarm of robots. Lect. Notes in Artifcial Intelligence 2801 (2003)
865–874
8. Martinoli, A., Easton, K., Agassounon, W.: Modeling swarm robotic systems:
A case study in collaborative distributed manipulation. Int. Journal of Robotics
Research 23(4) (2004) 415–436
9. SYMBRION: Project website (2009) http://www.symbrion.eu/.
10. REPLICATOR: Project website (2009) http://www.replicators.eu/.
11. Reynolds, C.W.: Flocks, herds, and schools. Computer Graphics 21(4) (1987)
25–34
12. Mataric, M.J.: Designing emergent behaviors: from local interactions to collec-
tive intelligence. Proc. of the Second Int. Conf. on From Animals to Animats 2:
simulation of adaptive behavior (1993) 432–441
13. Turgut, A., Çelikkanat, H., Gökçe, F., Şahin, E.: Self-organized flocking in mobile
robot swarms. Swarm Intelligence 2(2) (December 2008) 97–120
14. Hayes, A.T., Dormiani-Tabatabaei, P.: Self-organized flocking with agent fail-
ure: Off-line optimization and demonstration with real robots. In: Int. Conf. on
Robotics and Automation. (2002) 3900–3905
15. Balch, T., Hybinette, M.: Social potentials for scalable multi-robot formations.
Volume 1. (2000) 73–80 vol.1
16. Kelly, I.D., Keating, D.A.: Flocking by the fusion of sonar and active infrared
sensors on physical autonomous mobile robots. Proc. of the The Third Int. Conf.
on Mechatronics and Machine Vision in Practice 1 (1996) 1–4
17. Baldassarre, G., Nolfi, S., Parisi, D.: Evolving mobile robots able to display col-
lective behaviors. Artificial Life 9(3) (2003) 255–267
18. Bjerknes, J.D., Winfield, A., Melhuish, C.: An analysis of emergent taxis in a wire-
less connected swarm of mobile robots. In: IEEE Swarm Intelligence Symposium,
Los Alamitos, CA, IEEE Press (2007) 45–52
19. Partridge, B.L., Pitcher, T.J.: The sensory basis of fish schools: relative roles of
lateral line and vision. Journal of Comparative Physiology 135(4) (1980) 315–325
20. Wilensky, U.: Netlogo. Center for Connected Learning and Computer-Based Mod-
eling, Northwestern University. Evanston, IL (1999)
21. Moeslinger, C.: Video link (2009) http://zool33.uni-graz.at/artlife/flocking.
22. Swarmrobot: Project website (2009) http://www.swarmrobot.org/tiki-index.php.
23. ePuck: e-puck desktop mobile robot - website (2009) http://www.e-puck.org/.