Wall Climbing Emergent Behavior in a Swarm of Real-World Miniature

Autonomous Blimps

Tristan K. Schuler

1 a

, Cameron Kabacinski

, Daniel M. Lofaro

, Dhawal Bhanderi

Jennifer Nguyen

and Donald Sofge

1 b

Navy Center for Applied Research in AI (NCARAI), U.S. Naval Research Laboratory, Washington D.C, 20375, U.S.A.

Intern, U.S. Naval Research Laboratory, Washington D.C., 20375, U.S.A.

Keywords:

Miniature Autonomous Blimps, Swarm, Emergent Behavior.

Abstract:

Emergent behaviors in swarms can arise when agents with simple behavioral rules produce complex group

dynamics. In this work we develop both a physics-based simulation environment as well as a real-world

testbed for indoor miniature autonomous blimps to analyze sensor-driven emergent behavior. During the ﬂight

experiments, the blimps had global localization from an indoor motion capture system to follow waypoints

but used downward facing ultrasonic ping sensors for local-frame altitude control. After introducing a wall

to the environment and commanding the swarm to the other side of the unknown obstacle, a wall climbing

emergent behavior arose where the swarm agents climbed over each other as well as the wall to reach the goal.

We demonstrate how modifying the sensor characteristics between trials and changing the swarm size affects

this emergent behavior in both simulation and with real-world blimps, and compare the results.

1 INTRODUCTION

Emergent behavior can arise in swarms when an agent

with simple behavioral rules can produce complex

group dynamics. There are many instances of emer-

gent behavior arising in nature, especially with social

insects. One of the most popular emergent behav-

ior models is for the ﬂocking of birds or ﬁsh where

the initial conditions can lead to a common velocity

for the ﬂock (Cucker and Smale, 2007). Foraging

ants leave behind pheromones during foraging which

other ants follow and continue to reinforce, creat-

ing a “highway” of ant trafﬁc to the goal, typically

food (Bonabeau et al., 1997). Under certain con-

ditions, army ants will form “living bridges’ where

they will climb on top of one another to reach a goal

faster (Graham et al., 2017). Similarly, ﬁre ants self-

assemble to build towers as temporary shelters, espe-

cially after ﬂooding, a process that randomly emerges

(Phonekeo et al., 2017). Honeybees self-organize for

honeycomb creation and an overall pattern emerges

spontaneously from dynamic interactions among the

processes of depositing and removing brood, pollen

and honey (Camazine, 1991).

https://orcid.org/0000-0002-1567-322X

https://orcid.org/0000-0003-0153-3581

A common limitation to simulating swarm behav-

ior and algorithms is that there are frequently many

assumptions made such as only focusing on dynamic

behavior, having perfect communication, no catas-

trophic collisions, or even having agents with negligi-

ble size (Taylor and Nowzari, 2021). Several swarm

robot platforms and algorithm development efforts

have been inspired by nature. Researchers at Har-

vard developed the Bluebot underwater ﬁsh platform

and demonstrated common emergent behavior algo-

rithms on real agents, such as the ring state, and could

also sustain collisions with the tank and other agents

(Berlinger et al., 2021). Researchers at the Univer-

sity of Alabama explored waypoint navigation with

collidable micro-airships (Troub et al., 2017). The

GRASP lab at University of Pennsylvania developed

small UAVs that are capable of sustaining collisions

and then able to navigate without complete knowl-

edge of the environment, and then create a sparse

map of their surroundings (Mulgaonkar et al., 2017;

Mulgaonkar et al., 2020). Another group used sim-

ple robots with binary tactile sensors to determine if

a collision occurred and let the agents use this in-

formation for probabilistic localization (Mayya et al.,

2017; Mayya et al., 2018). Similarly, researchers

at ETH Zurich demonstrated navigation through an

unknown environment and low-level mapping by ex-

Schuler, T., Kabacinski, C., Lofaro, D., Bhanderi, D., Nguyen, J. and Sofge, D.

Wall Climbing Emergent Behavior in a Swarm of Real-World Miniature Autonomous Blimps.

DOI: 10.5220/0011636200003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 1, pages 225-232

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

225

ploiting collision information from an onboard IMU

(Lew et al., 2019). In this work, we try to bridge the

gap between simulated and real-world swarms using

an indoor autonomous blimp robotic platform.

Indoor lighter-than-air vehicles are a great robotic

platform for testing emergent behaviors on real-world

robots. One of the major beneﬁts to lighter-than-

air vehicles is that they can move in a 3D envi-

ronment with other agents without needing to ac-

count for the possibility of catastrophic collisions. In

this paper we use the Georgia Tech Miniature Au-

tonomous Blimps (GT-MAB) platform for produc-

ing wall-climbing emergent behavior on real-world

agents in a 3D space (Cho et al., 2017). In 2019,

researchers at the U.S. Naval Research Laboratory

demonstrated formation ﬂying with 25+ modiﬁed ver-

sions of the GT-MAB, the largest real-world demon-

stration of indoor autonomous blimps to date (Schuler

et al., 2019). Additional recent research with this plat-

form includes time synchronized path planning, in-

door localization, human interaction, and gesture con-

trol (Gibson et al., 2020; Yao et al., 2019; Yao et al.,

2017; Seguin et al., 2020; Srivastava et al., 2019). Re-

searchers at the U.S. Naval Research Laboratory also

explored swarm behavior in lighter-than-air vehicles

using universal global inputs from onboard sensing as

well as exploring visoelastic ﬂuid-inspired swarm be-

havior (Hall et al., 2021; McGuire et al., 2021).

In this paper, we set up an emergent behavior test

bed for miniature autonomous blimps in both simu-

lation and the real-world where the blimps climb an

unknown wall due to the inclusion of a downward

ultrasonic ping sensor for altitude control. We also

explore how adjusting the sensor characteristics or

parameters affects the emergent behavior. Addition-

ally, this paper discusses the development of a physics

based physics-based simulator capable of testing large

swarm sizes of indoor blimps. The simulator is vali-

dated by comparing the performance to real world ex-

periments with a smaller swarm of real-world agents.

2 WALL CLIMBING TESTBED

In our 2019 experiment using 25+ GT-MABs for for-

mation control, the agents used Vicon, an indoor in-

frared motion capture system, for horizontal waypoint

navigation. We installed downward-facing ultrasonic

ping sensors on the blimp gondola to do local-frame

altitude control. In the ﬂight demonstration we also

manually drove a Pioneer ground rover, shown In Fig-

ure 1, underneath the autonomous blimp formation

and demonstrated that individual agents rose after de-

tecting a local-frame altitude change from the Pioneer

Figure 1: 25+ blimp swarm experiment where wall climb-

ing phenomenon was discovered.

underneath. Occasionally, a blimp would drift away

from its target waypoint due to either HVAC currents,

hardware malfunctions, or occluded motion capture

markers. Some of these drifting agents would reach

a wall and start “climbing” it. The climbing phe-

nomenon arose from the downward facing ultrasonic

ping sensor detecting the wall and triggering an up-

ward velocity to reach the target altitude, which was

never obtained since the wall was in range. This in-

troduced the question of whether or not a group of

blimps with simple sensing could overcome obstacles

without path-planning. We hypothesized that a swarm

of blimps could climb a wall without knowledge of

the obstacle, and that the success rate of reaching the

other side of the wall could change by adjusting sen-

sor characteristics and the swarm size.

2.1 Agent Design

The lighter-than-air agents used in these experiments

are a modiﬁed version of the GT-MAB agents. The

GT-MABs are helium ﬁlled lighter-than-air vehicles

with a customized gondola attached to standard 36”

Mylar balloon envelopes. The gondola contains ﬁve

motor/propeller pairs for holonomic motion: two for

altitude control, two for forwards/backwards and yaw

control, and one for lateral control. Additionally,

the gondola also contains the motor controller, bat-

tery, and XBee radio for communicating with a host

computer. For indoor localization, the agents were

equipped with retro-reﬂective Vicon markers on the

tops of the envelopes. We lightly spray painted the

tops of the envelopes to reduce noisy reﬂections from

the metallic Mylar envelope.

The modiﬁed GT-MABs incorporate a downward

facing ultrasonic ping sensor for altitude control or

obstacle avoidance and can be seen in Figure 2. The

ultrasonic ping sensor implementation was designed

to be modular, so that different ping sensors, with

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

226

Figure 2: Modiﬁed Georgia Tech Miniature Autonomous

Blimps (GT-MAB) with downward facing ultrasonic sen-

sors for altitude control, and Vicon markers for localization.

Figure 3: MaxSonar Ultrasonic Ping Sensor beam width es-

timates for 30cm grid resolution and tested with a 3.5 inch

diameter dowel. Image 1 is narrowest beam width, and im-

age 2 is widest beam width available.

varying beam widths could easily be installed on

each. For these experiments we used two variations

of MaxSonar Ultrasonic ping sensors, the narrowest

and widest variations. Figure 3 shows the difference

between the narrow-beam and wide-beam sensors.

Both sensors have a maximum altitude range of 5.0 m

but signiﬁcantly different proximity cones for detect-

ing ultrasonic pings. For these experiments we use

the same controller architecture as our 25+ lighter-

than-air autonomous agents demonstration; horizon-

tal motion was controlled from motion-capture feed-

back and altitude was controlled from ultrasonic ping

feedback (Schuler et al., 2019).

2.2 Control Structure

Both the real-world and simulated blimps used a

control architecture of 4 independent proportional-

derivative (PD) feedback controllers for horizontal

position, yaw orientation, and altitude. The controller

PD outputs denoted as u

global

, u

global

, u

local

are calcu-

lated separately by:

= K

p j

+ K

i j

p j

′

+ K

d j

p j

(1)

where j = {x

global

, y

global

, z

local

}, e

p j

is the Eu-

clidean error, and K

, K

, and K

are the proportional,

integral and derivative gains.

A fourth PD controller for yaw about the global z

-axis is deﬁned by:

= K

+ K

dt + K

(2)

where u

is the resulting control rotational velocity

about the z-axis and e

is the error between the desired

orientation about z and the measured orientation. K

, and K

are the proportional, integral and deriva-

tive gains for the yaw controller.

The output of these 4 independent PD controllers

were velocities ˙x, ˙y, ˙z, and

[ ˙x, ˙y, ˙z,

] (3)

2.3 Simulation Environment

We developed a physics-based simulation environ-

ment to run experiments that would be challenging

to implement with real-world agents such as larger

swarm sizes and additional sensor characteristic dif-

ferences.

2.3.1 Simulation Architecture

The simulation architecture was developed in Cop-

peliaSim, previously known as VREP (Rohmer et al.,

2013). CoppeliaSim enables building dynamic robot

models out of simple and complex shapes that re-

spond to the environment, and allows individual agent

control through their own script. The simulator is

controlled through API calls to start/stop the simula-

tion and manipulate objects, which allows for simula-

tion automation and data collection.

2.3.2 Matching Simulated Blimps to GT-MAB

The simulated agents were matched to the physical

attributes of the real-world modiﬁed GT-MAB. Each

component, including the motors, envelope, gondola,

battery, and ballast, was measured and weighed, and

then recreated in simulation using simple shapes to

create a 1:1 scale model of the agent.

To make the simulation as realistic as possible,

real-world blimp characteristics were measured and

Wall Climbing Emergent Behavior in a Swarm of Real-World Miniature Autonomous Blimps

227

applied to the simulated agents. The most rele-

vant characteristics included measuring the maximum

force of a single motor, the blimp’s buoyancy, the

maximum velocity of the blimp, and approximate

angular dampening. Coppeliasim does not directly

support lift or aerodynamic simulation, so additional

forces are necessary to be applied to the model at each

simulation time step. Each agent’s control script di-

rectly interfaces with the simulated motors by apply-

ing a force at the motor locations and limited by the

measured maximum motor force.

To offset gravity and provide buoyancy, a con-

stant upward force is applied to the model’s center of

mass to estimate the lift force provided by the helium-

ﬁlled envelope. The exact force is based on the simu-

lated weight and a small ballast adjustment to provide

proper buoyancy (slightly below neutral). The buoy-

ancy characteristic was determined by dropping the

real blimp from a set height and recording the time

it took to fall, and then calibrating the ballast adjust-

ment in simulation.

To ensure a correct maximum achievable velocity,

a drag force was calculated and applied to the center

of mass at every simulation step. The drag force is

presented in Eq. 4 where F

is the drag force, ρ is the

air density, v is the velocity, c

is the drag coefﬁcient,

and A is the cross-section area. While each variable in

the drag force equation could be found independently,

by grouping ρ, A, c

, and the 1/2 factor together as β,

the drag force equation becomes simpliﬁed, as shown

in Eq. 5.

The blimp’s real world characteristics can be

leveraged to ﬁnd β as there are only two unknowns, F

and v, both of which can be found when the blimp is

moving with full forward throttle. The blimp’s max-

imum velocity happens when the blimp is no longer

able to accelerate and reaches a force balance, which

occurs when the two forward facing motors provide

a force equal to the drag force, as shown in Eq. 6,

where F

is the maximum force of a single motor.

The maximum velocity is measured by tracking the

blimp position over time. Thus, β will equal the max-

imum force provided by the forward motors divided

by the maximum velocity squared, shown in Eq. 7.

ρv

A (4)

β =

ρc

A, F

= βv

(5)

net

= ma, 2F

− F

= 0, F

= 2F

(6)

MAX

= β (7)

Figure 4: Real-world wall climbing testbed with 8 au-

tonomous blimps equipped with downward facing ultra-

sonic sensors for altitude control.

Once β was found, at each simulation step, the

blimp velocity could be calculated by the simulator

and multiplied by β to ﬁnd the drag force, as in Eq. 5,

and applied to the center of the blimp.

Additionally angular dampening was added to the

simulated agents to match real-world agents. With-

out any angular dampening, applied motor forces in

simulation caused a torque around the center of mass

which creates a rocking pendulum motion that is not

observed in the real world. The real blimp was held at

a 45-degree angle relative to the horizon and released

and timed until it settled back to horizontal equilib-

rium. Then, the simulated blimp was setup the same

and angular dampening coefﬁcients were tuned until

the settle time was the same between the real and sim-

ulated models.

Finally, a proximity sensor was added with match-

ing characteristics to the ultrasonic ping sensor in-

stalled on the real-world gondolas shown in Figure

3. A Gaussian noise distribution was added to the

sensor feedback to simulate noise from the real sen-

sor as well as inconsistencies with altitude holding

due to additional unknown forces such as unknown

air ﬂows or propeller wash. The same PD controller

structure from Section 2.2 was implemented on the

simulated agents for waypoint navigation with local

altitude sensing from the proximity sensor.

2.4 Experimental Setup

The simulation environment was setup with the same

parameters as the real-world experiment. A 2.5 m

wall was placed perpendicular to the desired path of

motion between the start and end way points. All

agents in the swarm were given the same starting and

ending waypoint in each trial. The agents had global

localization knowledge in the X-Y plane but used

a downward facing ultrasonic ping sensor for local-

frame altitude control. The agents then had 5 minutes

to attempt to maneuver to the other side of the un-

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

228

(1). (2). (3).

(4). (5). (6).

Figure 5: Fifty simulated autonomous blimps with a local altitude setpoint of 1.0 m and 20 s to reach starting waypoint after

initial random distribution (1-2). After 20 s the group waypoint is set to the opposite of the side of the wall; the swarm

attempts to move to the waypoint (3-5). The blimps start ”towering” with extended time at a target waypoint (6).

known wall. Figure 4 shows the real-world set up of

the experiment.

Figure 5 shows snapshots of a simulated wall

climbing trial with 50 agents at a 1.0 m local-frame

altitude setpoint. At the start of a trial, the agents were

randomly distributed on one side of the wall and had

a 20 s period to maneuver to the starting X-Y way-

point coordinate centered on the same side of the wall

as the swarm, shown in the ﬁrst two images of Fig-

ure 5. After 20 s, the waypoint changed to an X-Y

coordinate centered on the other side of the wall, and

the agents immediately try to maneuver to the new

waypoint. The agents were unaware of the wall and

their global altitude, only their global X-Y position

and local altitude readings from the downward facing

proximity sensors. An emergent behavior arose where

the blimps “climbed” the wall as well as other nearby

agents while trying to maneuver to the new waypoint,

shown in images 3-6 in Figure 5. By changing sensor

characteristics such as the altitude set point and beam

width, as well as the swarm size, the success rate of

the wall climbing emergent behavior changed.

Additionally, a secondary emergent behavior of

”towering” occasionally occurred at the starting or

ending waypoints when the swarm has extended time

at the waypoint. Figure 6 shows the towering behav-

ior in simulated and real world agents. Towering oc-

curs when an agent ﬂoats above another and triggers

a new local-frame altitude reading from the ultrasonic

sensor. The balloons do not start in a towering forma-

tion, but can maneuver on top of other agents for sev-

eral reasons including collisions from other agents,

prop wash from other agents, and sensor noise caus-

ing oscillating altitude control. Towering can happen

Figure 6: Towering emergent behavior with an 8 agent

swarm in both real-world and simulated blimps.

with as little as two agents, but is more prominent

with larger swarms.

Multiple trials were simulated and conducted with

physical agents with varying sensor characteristics in-

cluding beam width and altitude set point, as well as

varying the size of the swarm to observe effects on

the wall-climbing emergent behavior. Wall climbing

success was binary for individual agents; an individ-

ual success was denoted by crossing the plane of the

wall. The overall success rate of the swarm was then

the average percentage of agents that made it to the

other side of the wall. Five (5) trials were conducted

for each set of independent variables with the swarm

Wall Climbing Emergent Behavior in a Swarm of Real-World Miniature Autonomous Blimps

229

Figure 7: Simulated wall climbing success rates for downward facing narrow-beam-width proximity sensor (top) and wide-

beam-width (bottom) sensor for varying altitude setpoints and swarm sizes (color gradients).

wall climbing success rates averaged. The indepen-

dent variables in each trial included local-frame al-

titude setpoint from proximity sensor (0.5m, 1.0m,

1.5m, and 2.0m), proximity sensor beamwidth (nar-

row, wide), and swarm size (1, 5, 8, 10, 20, 50, and

100 agents).

3 RESULTS & DISCUSSION

Figure 7 shows the simulated swarm’s average wall

climbing success rates (along y-axis) after 5 trials for

3 varying parameters: proximity sensor beam width

(narrow on top, wide on bottom), swarm size (col-

ored gradient), and the altitude set point (along x-

axis). As hypothesized, both a larger swarm size and

higher altitude set points increased the overall swarm

wall-climbing success rate. The 2.0 m altitude set-

point resulted in a 100% wall climbing success rate

for any number of simulated agents for both the wide

and narrow proximity sensor. Surprisingly, the wide-

angle proximity sensor had minimal effect on the wall

climbing success rate, both on simulated agents and

the real-world agents. On average, the wide-angle

ultrasonic sensor produced a 5-10% increase in wall

climbing success rate compared with the narrow beam

width sensor. This gap in performance between the

narrow and wide beam width sensors also decreased

the larger the swarm size was. Therefore, like the real-

world experiments, the altitude set point was the driv-

Figure 8: Wall climbing success rate for real-world blimps

equipped with wide angle ultrasonic ping sensors at various

altitude setpoints.

ing force for wall climbing emergent behavior, not the

beamwidth of the sensor.

Similar trials were conducted with a swarm of

8 real-world blimps. Figure 8 shows the individual

trials of the 8-agent swarm. For the most part, the

agents that successfully crossed the wall were ran-

dom between trials; however, there were exceptions.

Throughout the trials, blimp #7 never crossed the wall

although it was fully operational. Similarly, in the 2.0

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

230

Figure 9: Experimental and simulated wall climbing suc-

cess rate for 8 real-world blimps at various altitude setpoints

and ping sensor beam widths.

m experiments blimps #2, #6, and #11 made it across

the wall every time.

Figure 9 shows the wall climbing success rate dif-

ference between simulation and the real-world exper-

iment for an 8-agent swarm. The overall trend was

similar, with a higher altitude setpoint resulting in a

higher wall climbing success rate for both simulated

and real-world agents. However, the simulated en-

vironment had higher success rates for each altitude

setpoint. The 2.0 m setpoint had a 100% success rate

for simulated agents whereas only 65% of the swarm

made it across on average with the real-world blimps.

It is important to note that the simulated agents

perform better than the real-world agents, for many

reasons. While efforts were made to match the sim-

ulation to the dynamics of the real-world blimps, as

discussed in section 2.3.2 there’s still many unknown

or hard to simulate variables. A major inconsis-

tency that’s challenging to simulate is random hard-

ware malfunctions or failures, which happened fre-

quently. The most common hardware issues were

typically motors with weakened thrust proﬁles after

frequent use or communication issues with either the

mission control computer, or the motion capture sys-

tem. These hardware issues were a major reason why

we felt the need to develop a MAB simulator to test

larger swarms of MABs without constantly making

robot repairs or solving the hard problem of indoor

global localization with a large number of agents.

The gray numbered blimps in Figure 8 shows

blimps that were not included in the trials. These

were typically hardware issues that could be repaired

or ﬁxed in between trials, but sometimes the issue was

a permanent failure that required hardware replace-

ment. The required ballast and buoyancy force varies

from blimp to blimp due to helium leakage and re-

quires occasional reﬁlling before experiments. Ad-

ditionally, small changes in temperature as well as in-

ternal drafts from the HVAC system or other miscella-

neous indoor currents can also have major effects on

blimp performance; while these two variables could

be simulated, they are hard to accurately match to the

real-world testbed.

4 CONCLUSIONS AND FUTURE

WORK

In this work we develop a simulation and experimen-

tal environment for exploring emergent behaviors in

swarms of indoor miniature autonomous blimps. We

demonstrate a wall climbing emergent behavior in

a swarm of autonomous blimps experimentally and

in simulation through changing the characteristics of

a very simple ultrasonic ping sensor and the swarm

size. As the local-frame altitude set point increases,

the success rate of wall climbing increases; also, as

the swarm size increases, the wall climbing success

rate increases. With extended time at a group way-

point, a secondary emergent behavior of ”towering”

would also occasionally occur.

In the future we would like to change the con-

trol scheme to explicitly include potential ﬁelds to

see if the emergent behavior changes, as well as add

more known and unknown obstacles. In these experi-

ments the individual agents’ behavior mimics 2D po-

tential ﬁelds in an unknown environment. There is

an attractive force in the XY plane attracting all of

the agents to the other side of the room and a re-

pulsive force along the Z-axis from the ﬂoor, wall,

and other agents all acting as repulsive forces. Typ-

ically, the wall acts as a local maximum for individ-

ual agents with a low altitude setpoint (<1.5). How-

ever, with a large enough swarm, the attractive force

can overcome the cumulative repulsive forces from

other agents when they all reach the wall. We would

also like to test the emergent behaviors with differ-

ent styles of indoor autonomous blimp platforms that

are currently used in the research community such as

ﬂapping wing blimps, cigar-shaped and ﬁsh-inspired

blimps, and spherical blimps (Lin et al., 2022). Dif-

ferent shapes and sizes of blimps will introduce vary-

ing motion proﬁles, agility, controllability, and addi-

tional sensing capability on larger blimps.

ACKNOWLEDGMENTS

This work was performed at the U.S. Naval Research

Laboratory for the project “Adaptive Real-Time Al-

gorithms for Multiagent Cooperation in Adversarial

Environments”. The views, positions and conclusions

Wall Climbing Emergent Behavior in a Swarm of Real-World Miniature Autonomous Blimps

231

expressed herein reﬂect only the authors’ opinions

and expressly do not reﬂect those of the U.S. Naval

Research Laboratory, nor those of the Ofﬁce of Naval

Research. The authors would like to thank the Geor-

gia Tech Systems Research Lab for providing hard-

ware for the modiﬁed GT-MABs. We would like to

thank several additional U.S. Naval Research Lab in-

terns for assisting with the real-world wall climbing

experiments, including Alex Maxseiner, Brian Mate-

vich, Divya Srivastav, Tony Lin, and Richard Hall.

REFERENCES

Berlinger, F., Gauci, M., and Nagpal, R. (2021). Implicit

coordination for 3d underwater collective behaviors in

a ﬁsh-inspired robot swarm. Science Robotics, 6(50).

Bonabeau, E., Theraulaz, G., Deneubourg, J.-L., Aron, S.,

and Camazine, S. (1997). Self-organization in social

insects. Trends in ecology & evolution, 12(5):188–

193.

Camazine, S. (1991). Self-organizing pattern formation on

the combs of honey bee colonies. Behavioral ecology

and sociobiology, 28(1):61–76.

Cho, S., Mishra, V., Tao, Q., Vamell, P., King-Smith, M.,

Muni, A., Smallwood, W., and Zhang, F. (2017). Au-

topilot design for a class of miniature autonomous

blimps. In 2017 IEEE conference on control technol-

ogy and applications (CCTA), pages 841–846. IEEE.

Cucker, F. and Smale, S. (2007). Emergent behavior

in ﬂocks. IEEE Transactions on automatic control,

52(5):852–862.

Gibson, J., Schuler, T., McGuire, L., Lofaro, D. M., and

Sofge, D. (2020). Swarm and multi-agent time-based

a* path planning for lighter-than-air systems. Un-

manned Systems, 8(03):253–260.

Graham, J. M., Kao, A. B., Wilhelm, D. A., and Garnier,

S. (2017). Optimal construction of army ant living

bridges. Journal of theoretical biology, 435:184–198.

Hall, R., Maxseiner, A., Matejevich, B., Sofge, D. A., and

Lofaro, D. M. (2021). Emergent behavior in swarms

with on-board sensing. In 2021 IEEE Symposium Se-

ries on Computational Intelligence (SSCI). IEEE.

Lew, T., Emmei, T., Fan, D. D., Bartlett, T., Santamaria-

Navarro, A., Thakker, R., and Agha-mohammadi, A.-

a. (2019). Contact inertial odometry: collisions are

your friends. In The International Symposium of

Robotics Research, pages 938–958. Springer.

Lin, T. X., Rossouw, M., Maxseiner, A. B., Schuler, T., Gar-

ratt, M. A., Ravi, S., Zhang, F., Lofaro, D. M., and

Sofge, D. A. (2022). Miniature autonomous blimps

for indoor applications. In AIAA SCITECH 2022 Fo-

rum, page 1834.

Mayya, S., Pierpaoli, P., Nair, G., and Egerstedt, M. (2018).

Localization in densely packed swarms using inter-

robot collisions as a sensing modality. IEEE Trans-

actions on Robotics, 35(1):21–34.

Mayya, S., Pierpaoli, P., Nair, G. N., and Egerstedt, M.

(2017). Collisions as information sources in densely

packed multi-robot systems under mean-ﬁeld approx-

imations. In Robotics: Science and Systems, vol-

ume 13.

McGuire, L., Schuler, T., Otte, M., and Sofge, D. (2021).

Viscoelastic ﬂuid-inspired swarm behavior to reduce

susceptibility to local minima: The chain siphon al-

gorithm. IEEE Robotics and Automation Letters,

7(2):1000–1007.

Mulgaonkar, Y., Liu, W., Thakur, D., Daniilidis, K., Tay-

lor, C. J., and Kumar, V. (2020). The tiercel: A novel

autonomous micro aerial vehicle that can map the en-

vironment by ﬂying into obstacles. In 2020 IEEE In-

ternational Conference on Robotics and Automation

(ICRA), pages 7448–7454. IEEE.

Mulgaonkar, Y., Makineni, A., Guerrero-Bonilla, L., and

Kumar, V. (2017). Robust aerial robot swarms without

collision avoidance. IEEE Robotics and Automation

Letters, 3(1):596–603.

Phonekeo, S., Mlot, N., Monaenkova, D., Hu, D. L., and

Tovey, C. (2017). Fire ants perpetually rebuild sinking

towers. Royal Society open science, 4(7):170475.

Rohmer, E., Singh, S. P., and Freese, M. (2013). V-rep:

A versatile and scalable robot simulation framework.

In 2013 IEEE/RSJ International Conference on Intel-

ligent Robots and Systems, pages 1321–1326. IEEE.

Schuler, T., Lofaro, D., McGuire, L., Schroer, A., Lin, T.,

and Sofge, D. (2019). A study of robotic swarms

and emergent behaviors using 25+ real-world lighter-

than-air autonomous agents (lta3). In 2019 3rd In-

ternational Symposium on Swarm Behavior and Bio-

Inspired Robotics (SWARM).

Seguin, L., Zheng, J., Li, A., Tao, Q., and Zhang, F. (2020).

A deep learning approach to localization for naviga-

tion on a miniature autonomous blimp. In 2020 IEEE

16th International Conference on Control & Automa-

tion (ICCA), pages 1130–1136. IEEE.

Srivastava, D., Lofaro, D., Schuler, T., and Sofge, D.

(2019). Gesture-based interface for multi-agent and

swarm formation control. In 2019 3rd International

Symposium on Swarm Behavior and Bio-Inspired

Robotics (SWARM).

Taylor, C. and Nowzari, C. (2021). The impact of catas-

trophic collisions and collision avoidance on a swarm-

ing behavior. Robotics and Autonomous Systems,

140:103754.

Troub, B., DePineuil, B., and Montalvo, C. (2017). Simula-

tion analysis of a collision-tolerant micro-airship ﬂeet.

Int’l Journal of Micro Air Vehicles, 9(4):297–305.

Yao, N., Anaya, E., Tao, Q., Cho, S., Zheng, H., and Zhang,

F. (2017). Monocular vision-based human follow-

ing on miniature robotic blimp. In 2017 IEEE In-

ternational Conference on Robotics and Automation

(ICRA), pages 3244–3249. IEEE.

Yao, N.-s., Tao, Q.-y., Liu, W.-y., Liu, Z., Tian, Y., Wang,

P.-y., Li, T., and Zhang, F. (2019). Autonomous ﬂy-

ing blimp interaction with human in an indoor space.

Frontiers of Information Technology & Electronic En-

gineering, 20(1):45–59.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

232