ArticlePDF Available


Invasive alien species threaten biodiversity worldwide and contribute to biotic homogenization, especially in freshwaters, where the ability of native animals to disperse is limited. Robotics may offer a promising tool to address this compelling problem, but whether and how invasive species can be negatively affected by robotic stimuli is an open question. Here, we explore the possibility of modulating behavioural and life-history responses of mosquitofish by varying the degree of biomimicry of a robotic predator, whose appearance and locomotion are inspired by natural mosquitofish predators. Our results support the prediction that real-time interactions at varying swimming speeds evoke a more robust antipredator response in mosquitofish than simpler movement patterns by the robot, especially in individuals with better body conditions that are less prone to take risks. Through an information-theoretic analysis of animal-robot interactions, we offer evidence in favour of a causal link between the motion of the robotic predator and a fish antipredator response. Remarkably, we observe that even a brief exposure to the robotic predator of 15 min per week is sufficient to erode energy reserves and compromise the body condition of mosquito-fish, opening the door for future endeavours to control mosquitofish in the wild.
Cite this article: Polverino G, Karakaya M,
Spinello C, Soman VR, Porfiri M. 2019
Behavioural and life-history responses of
mosquitofish to biologically inspired and
interactive robotic predators. J. R. Soc. Interface
16: 20190359.
Received: 24 May 2019
Accepted: 7 August 2019
Subject Category:
Life SciencesEngineering interface
Subject Areas:
biomimetics, evolution, environmental science
animal personality, bioengineering,
biomimetics, body condition, invasive species,
predation risk
Author for correspondence:
Maurizio Porfiri
Electronic supplementary material is available
online at
Behavioural and life-history responses of
mosquitofish to biologically inspired and
interactive robotic predators
Giovanni Polverino1,2, Mert Karakaya3, Chiara Spinello3, Vrishin R. Soman3
and Maurizio Porfiri3,4
Centre for Evolutionary Biology, University of Western Australia, Perth, Australia
Department of Biology and Ecology of Fishes, Leibniz-Institute of Freshwater Ecology and Inland Fisheries,
Berlin, Germany
Department of Mechanical and Aerospace Engineering, and
Department of Biomedical Engineering,
Tandon School of Engineering, New York University, Brooklyn, NY, USA
MP, 0000-0002-1480-3539
Invasive alien species threaten biodiversity worldwide and contribute
to biotic homogenization, especially in freshwaters, where the ability of
native animals to disperse is limited. Robotics may offer a promising tool
to address this compelling problem, but whether and how invasive species
can be negatively affected by robotic stimuli is an open question. Here, we
explore the possibility of modulating behavioural and life-history responses
of mosquitofish by varying the degree of biomimicry of a robotic predator,
whose appearance and locomotion are inspired by natural mosquitofish
predators. Our results support the prediction that real-time interactions
at varying swimming speeds evoke a more robust antipredator response
in mosquitofish than simpler movement patterns by the robot, especially
in individuals with better body conditions that are less prone to take risks.
Through an information-theoretic analysis of animalrobot interactions, we
offer evidence in favour of a causal link between the motion of the robotic
predator and a fish antipredator response. Remarkably, we observe that
even a brief exposure to the robotic predator of 15 min per week is sufficient
to erode energy reserves and compromise the body condition of mosquito-
fish, opening the door for future endeavours to control mosquitofish in
the wild.
1. Introduction
The presence of animal species in areas where they are not native is common
across the globe, with tremendous costs for both human activities and the
ecological integrity of those areas [1,2]. Despite efforts from both governmental
and academic institutions, existing methods for eradicating invasive alien
species (IAS) or mitigating their negative effects remain labour-intensive,
economically unviable, and, often, ineffective [3].
Freshwater animals are particularly vulnerable to IAS, whereby native
species are confined to smaller water bodies and their ability to disperse is
limited compared to other ecosystems [4]. Mosquitofish (Gambusia affinis,
Baird and Girard, and Gambusia holbrooki, Girard) are among the most widely
diffused freshwater IAS in the globe, and their negative impact on indigenous
animal communities (via aggressive behaviours and/or predation [58]) has
been recognized by the International Union for Conservation of Nature that
listed mosquitofish among the worlds hundred worst IAS [9].
Technical efforts to eradicate mosquitofish from water bodies and mitigate
their negative impact on the native fauna are, however, limited. For example,
increasing the structural complexity of the environment through artificial
refugia was successful in reducing mortality in barrens topminnow (Fundulus
julisia, Williams and Etnier) exposed to mosquitofish under laboratory settings,
© 2019 The Author(s) Published by the Royal Society. All rights reserved.
but beneficial effects from artificial refugia disappeared in the
wild [10]. Similarly, the use of fish toxicants to combat
the spread of invasive mosquitofish resulted in detrimental
consequences for native fish [11]. The utilization of floating
traps to target mosquitofish near the water surface has
been shown to be a successful technique, but it is a labour-
intensive process that can be pursued only in small sites
and for short periods of time [12].
Robotics constitutes a promising tool for addressing some
of these challenges, by offering a versatile, customizable, and
consistent approach to modulate the behavioural response of
live animals [1315]. Particularly relevant are experiments
that have shown the possibility of eliciting behavioural
responses in freshwater fish through biologically inspired
robots triggering a costbenefit decision process [1621].
The use of robotics in the study of predatorprey interactions
might afford the design of new hypotheses-driven studies
that could unfold the basis of fear and anxiety in prey fish
[2225] and illuminate the evolutionary consequences of non-
lethal exposure to predators [26,27]. Just as robotics might
bring new scientific insight into predatorprey interactions,
it also contributes to ethics in animal experimentation by
minimizing potential harm to live animals.
In particular, previous research efforts from our group
indicate that a robotic fish can be designed to repel mosquito-
fish [28] and simultaneously attract non-invasive fish under
laboratory settings [19]. The possibility to isolate fish from
one species to another allows safeguarding non-invasive
species from the aggressive attitudes of mosquitofish, thereby
providing compelling evidence for the use of biologically
inspired robots as a possible method for the selective control
of mosquitofish in the wild. However, the technology to
deploy autonomous robotic fish in a complex ecological
environment to control the behaviour of mosquitofish is
still in its infancy, calling for a scientifically principled under-
standing of how mosquitofish interact with biologically
inspired robotic stimuli.
Mosquitofish can adjust their behavioural and life-history
strategies in response to varying environmental conditions,
especially in the attempt to minimize risks of predation [29].
Mosquitofish are typically less prone to take risks [30] and
invest less in reproduction [31] and energy reserves [32]
under predation risk than in more beneficial conditions, with
plastic adjustments associated with predation risks that
can eventually result in the whole body morphology of
mosquitofish to be reshaped [33]. Visual cues represent
the predominant factor for predator recognition in most fresh-
water fish [34], especially mosquitofish [28,35], and a growing
literature has provided convincing evidence that visual cues
from animated images [3638] and biologically inspired
robots [19,28] can be used to influence mosquitofish behaviour.
While experiments comparing mosquitofish behavioural
response to computer-animated and robotic stimuli are pre-
sently lacking, evidence from other freshwater fish suggests
that visual stimuli associated with a biologically inspired
robotic predator might elicit a stronger response than compu-
ter-animated images [23]. Experiments in [23] have compared
the fear response of zebrafish (Danio rerio, Hamilton) evoked
by live predator fish, a robotic replica of the predator fish,
and computer-animated images of the predator fish,
determining that: (a) the robot caused a robust avoidance
response in zebrafish that was comparable to that observed
for live predators, while computer-animated images did
not, and (b) individual responses were more consistent over
time when zebrafish were exposed to the robot than to
live predators and computer-animated images. In addition
to these methodological observations, practical consider-
ations towards future deployment in the wild favour the
use of robots over computer-animated images. In fact, practi-
cality challenges the use of computer-animated images in the
wild, where it may be unfeasible to employ computer screens
or projectors. Based on these methodological and practical
aspects, we favour the use of robotic stimuli in place of
computer-animated images.
In this study, we sought to test whether behavioural and
life-history responses of mosquitofish could be modulated
through a robotic predator whose visual appearance and
locomotion were inspired by mosquitofish predators, the
largemouth bass (Micropterus salmoides, Lacépède; figure 1).
Largemouth bass coexist with mosquitofish in the wild
1 cm
1 cm
Figure 1. Schematic for (a) the overview of experimental apparatus, (b) the three-dimensional representation of the biologically inspired predator replica, (c) the
biologically inspired predator replica used for experiments and (d) a picture of a juvenile largemouth bass. (Online version in colour.) J. R. Soc. Interface 16: 20190359
and constitute their most common predators [39,40], with
mosquitofish representing over 80% of the fish consumed
by juvenile largemouth bass in their native environments
[41]. Our biologically inspired robotic predator was designed
to take advantage of the innate antipredator behaviour that
largemouth bass induce in mosquitofish under laboratory
settings [31,42]. We repeatedly exposed mosquitofish to
robotic predators varying in their degree of biomimicry to
disentangle the relative contribution of the robot swimming
and its interactivity on both behavioural and life-history
adjustments associated with antipredator responses in
mosquitofish. We hypothesized that: (a) visual stimuli from
the robotic predator would repel mosquitofish, as suggested
in [28,36], (b) increasing the degree of biomimicry in the
motion of the robot would increase antipredator behaviours
and impact life-history strategies (i.e. energy reserves) in
mosquitofish, and (c) individuals would differ from each
other in the extent of their antipredator responses [43], with
individuals with high future expectations (i.e. individuals
with high energy reserves) being more risk-averse than
others [44].
From a methodological point of view, our study contri-
butes to the state of the art in animalrobot interactions
[13,14,45] along several research directions. First, we estab-
lished a robotic platform that allows for tailoring the degree
of complexity of the interaction through a closed-loop control
system, integrating real-time tracking and high-precision
robotics. Through this platform, we successfully varied the
degree of biomimicry of the interactive robotic predator,
by simulating random attacks towards the fish at either con-
stant or increasing speed. This experimental manipulation
effectively allowed for the quantification of the relative
contributions of typical locomotory patterns of predators in
triggering antipredator responses in mosquitofish. Second,
we shied away from a rigid prototype, in favour of a soft
robotic replica that incorporates a spine-like structure to
promote natural oscillations that are reminiscent of body
undulations, which are known to be critical for fishrobot inter-
actions in the water [18,21]. Third, we integrated traditional
means of behavioural analysis with modern elements of
dynamical systems theory, through the information-theoretic
frameworkof transfer entropy [46]. Through the lens of transfer
entropy, we demonstrated an improved comprehension of the
antipredator response of mosquitofish, by testing for potential
cause-and-effect relationships between the motion of the
robotic predator and mosquitofish antipredator response.
Finally, although few recent studies have considered behav-
ioural response of animals repeatedly confronted with robots
[47,48], a detailed study of individual variation in mosquitofish
behaviour was lacking, especially in the context of life-history
consequences of the exposure to robotic stimuli.
Although focused on mosquitofish, the theoretical and
methodological underpinnings of this work could inform
research on other IAS, whose presence in the environment
is also a threat to biodiversity and economy. For example,
recent studies have demonstrated the possibility of using
robots inspired by live predators to influence the behaviour
of locusts (Locusta migratoria, Linnaeus), a major pest for
human agricultural economies and ecosystems stability
[49,50]. Similarly, the peregrine falcon-like robot Robird
has been recently presented for deployment in the aviation
industry to deter birds from flying in the vicinity of
aircrafts [51].
2. Material and methods
2.1. Study organism and maintenance
A total of 150 wild-caught western mosquitofish (Gambusia
affinis, Baird and Girard) were purchased from a commercial
supplier (Carolina Biological Supply Co., Burlington, NC, USA)
and were acclimatized for 1 day in stock tanks. Then, 75 focal
individuals (average body length of 2.9 ± 0.3 cm) were randomly
selected from stock tanks, with sick individuals and fish showing
physical and/or behavioural anomalies excluded a priori.
Focal fish were housed individually in transparent Plexiglas
cylinders (10 cm diameter), placed within a large housing tank
(185 × 47 × 60 cm, length, width and height) and submerged
in water for 10 cm, as in [29,52]. The lateral surface of the
transparent cylinders was perforated to promote water circula-
tion across separate cylinders, affording visual and chemical
interaction among individuals despite physical isolation. This
housing scheme prevented aggression, competition for resources,
and sexual harassment among mosquitofish, with each cylinder
marked with a unique identification code to facilitate the
identification of individuals over time. The position of the cylin-
ders was periodically randomized to allow visual and chemical
interactions among all fish. Fish were acclimatized in the cylin-
ders for one month before experiments, and they were housed
in these cylinders for the whole duration of the study (approx.
three months).
Fish were kept under a 12 h light/12 h dark photoperiod and
fed with commercial flake food (Nutrafin max; Hagen Corp.,
Mansfield, MA, USA) once a day. Water parameters were
checked daily, with temperature and pH maintained at 26°C
and 7.2, respectively, throughout the study.
2.2. Experimental set-up
2.2.1. Experimental arena for behavioural tests
Behavioural trials were performed in an experimental arena
(44 × 30 × 30 cm, length, width and height), filled with 10 cm of
conditioned water (figure 1a). The walls and the bottom surface
of the arena were covered with white opaque contact paper to
control for external disturbance and optimize automated compu-
ter tracking of fish motion during trials. Two 38 W fluorescent
tubes (All-Glass Aquarium, UK) were mounted 130 cm above
ground and were used to provide homogeneous illumination
to the apparatus. A high-resolution webcam (Logitech C920
webcam, Lausanne, Switzerland) was mounted 140 cm above
the floor for a complete overview of the experimental arena.
2.2.2. Robotic platform and predator replica
The experimental arena was supported by aluminium T-slotted
bars 29 cm above the ground to allow the placement of the
robotic platform underneath (figure 1a). The platform allowed
for manoeuvring the robotic replica along the 3 d.f.: 2 d.f. were
controlled for in-plane translational motion of the replica and
1 d.f. served to adjust the predator body rotation. The replica
was magnetically connected to the platform through a 3D-
printed base made of polylactic acid filaments (3.2 cm ×
1.0 cm × 0.6 cm length, width and height) containing two circular
neodymium magnets (0.63 cm thick and 0.3 cm diameter) and an
acrylic rod (4 cm length and 0.62 cm diameter; figure 1ac). The
in-plane translational motion was based on a Cartesian plotter
(XY Plotter Robot Kit, Makeblock Co. Ltd, Shenzhen, China)
and the body rotation was controlled via a stepper motor
(NEMA 14, Pololu Corp., Las Vegas, NV, USA). Further details
on the robotic platform are in the electronic supplementary
material. The platform was originally designed in [53] to study
zebrafish social behaviour and used in [54] to examine zebrafish
learning. J. R. Soc. Interface 16: 20190359
Locomotory patterns of the predator replica were inspired
by pilot tests performed on three juvenile largemouth bass
(7.0 ± 0.5 cm), purchased from Teichwirtschaften Armin Kittner
in Quitzdorf am See, Germany (https://www.teichwirtschaft-, before the beginning of the experiment (figure 1d).
Live bass were placed individually in the experimental arena
and their behaviour was recorded over 30 min. Swimming
trajectories and swimming speeds were then obtained through
an offline tracking software developed by our group [55].
Mean and maximum swimming speed measured in the pilot
tests and a swimming trajectory representative of the bass behav-
iour in the experimental arena were used for the motion of the
predator replica.
The morphology and coloration of the replica were also
chosen to capture salient features of juvenile largemouth bass
(figure 1bd). Towards this aim, we took photos of the live
bass from different angles and estimated their body dimensions
using a dedicated software (ImageJ, National Institute of Health,
Bethesda, Maryland, USA). The body morphology of the replica
was accordingly modelled in Solidworks (Dassault Systèmes
SolidWorks Corp., Waltham, Massachusetts, USA) to create a
three-dimensional design and, then, a solid mould. A spine-like
structure in polylactic acid filament material was 3D-printed
and integrated within the 3D-printed mould of the predator
replica together with two glass eyes, relatively smaller than in
live bass (figure 1b). Then, the mould was filled with non-toxic
and aquarium safe silicone (Dragon Skin 10 Medium, Smooth-
On, Macungie, PA, USA) and let dry. The spine-like structure
provided support to the weight of the silicone body of the replica
and facilitated body oscillations during swimming. Lastly, the
silicone body of the replica was hand-painted using non-toxic,
aquarium safe and silicone-based light grey and silver paints
(Smooth-On, Inc., Macungie, PA, USA) to mimic the character-
istic coloration pattern of largemouth bass (figure 1c). Colour
reflectance comparisons between live bass and its robotic replica
were not performed. However, non-toxic pigments used to paint
the body of the robotic replica have been shown to be effectively
perceived as natural pigments in bluefin killifish (Lucania goodie,
Jordan) [20], a freshwater fish with well-developed vision like
The moulded silicone body with glass eyes and spine-like
structure was attached to a clear acrylic rod, connected to
3D-printed base with magnets. The clear acrylic rod allowed
for setting the swimming depth of the biologically inspired
predator replica in the middle of the water column, i.e. where
the antipredator response of mosquitofish is known to be the
strongest [28].
2.2.3. Experimental conditions and live tracking
We designed a series of experimental conditions with robotic
replicas varying their motion to proxy different degrees of
biomimicry of live predators. In one control condition, the
experimental fish were tested in the absence of the replica (no
predator, NP). In a second control condition, the replica was
motionless and positioned randomly within the arena before
each trial started ( predator motionless, PM). In the four exper-
imental conditions where a swimming replica was employed,
the replica swam on either the predetermined trajectory inspired
by live bass (open-loop, OL) or it alternated between the
predetermined trajectory and targeted real-time interactions
(closed-loop, CL) with the focal fish. In two OL conditions, the
biologically inspired predator replica followed the predeter-
mined swimming trajectory, either at a varying speed based on
the motion of the live predator (OL1) or at a constant speed
(OL2). In condition OL2, the trajectory from the live bass
was processed to manoeuvre the replica at a constant speed.
Specifically, we locally fitted the trajectory using cubic splines
(interparc, Copyright (c) 2012 John DErrico) and placed equally
spaced waypoints on the splines such that the replica would
move at a constant speed. The constant speed was chosen to
be 6 cm s
to match the mean speed observed in juvenile
largemouth bass in our pilot tests and provide a dynamically
rich visual stimulation for mosquitofish. The same speed was
used as the mean value of the speed profile in condition OL1,
consistently scaling experimental observations.
In the CL conditions, the replica, besides following swimming
trajectories at a varying speed, was programmed to interact in real
time with the focal fish and to perform simulated attacks at
random. However, the replica always performed an attack every
minute of the trial for a total of 15 attacks. During an attack, the
replica either accelerated to attain a large speed (20 cm s
; CL1)
comparable to the maximum speed of live bass attacking a prey
[56], or swam at a constant speed towards the fish (6 cm s
; CL2).
When the replica was commanded to attack the focal fish, its
motion was a function of the distance from the fish. For CL1 con-
dition, if the distance between the fish and replica was less than
1 cm, the replica would only change its heading towards the
direction of the focal fish and return to the original heading;
for distances between 1 and 10 cm (inspection zone in [57]), the
replica would change its heading, accelerate towards the fish at
20 cm s
, and stop at approximately 1 cm from it; and for dis-
tances larger than 10 cm, the replica would change the
heading, accelerate at 20 cm s
until reaching a speed of
20 cm s
, and maintain this speed until stopping at 1 cm from
the fish. For CL2 condition, if the distance between the fish
and replica was less than 1 cm, the replica would only change
its heading towards the direction of the focal fish and return to
the original heading. For any distance greater than 1 cm, the
replica would change its heading, and attack the fish with a
constant speed of 6 cm s
and stop at 1 cm from the fish.
After an attack was completed, the replica returned to its
original position prior to the attack and restarted swimming
along the predetermined trajectory until the next attack. Notably,
the region in which the robotic replica swam was smaller than
the actual size of the experimental arena to allow at least 1 cm
from the extremities of the replicas body (i.e. head and caudal
fin) and the edges of the arena. This tolerance permitted
smooth operation of the robotic platform and avoided collision
with the walls of the arena. Further details on the real-time track-
ing system implemented for CL conditions are in the electronic
supplementary material.
The custom-made software was calibrated on the exact size of
an individual fish at each trial (week) separately and used to
calculate the following quantities: distance moved (cm), time
spent freezing (s), speed variance during swimming (cm
mean distance from the predator replica (cm), predator inspec-
tion (counts) and time spent within one-body length from the
wall (s)i.e. thigmotaxis [58]. In particular, if a fish moved at a
speed less than half of its body length per second for two
consecutive seconds, it was considered as freezing [59]. Predator
inspection was estimated according to standard protocols
developed for guppies (Poecilia reticulata, Peters) [57], a poeciliid
species closely related to mosquitofish. In particular, we counted
the number of events that a fish voluntarily approached the
predator replica by entering the 10 cm region around the replica
while actively swimming in its direction, i.e. at an angle lower
than ± 90° from the replicas head [57]. The distance from the
wall used to estimate thigmotaxis was selected based on pilot
tests in which mosquitofish were exposed to the same robotic
predator replica used in this study. Details of data extraction
and tracking system are in the electronic supplementary material.
Notably, reduced activity (in the form of short travelled
distances and prolonged freezing) and large number of predator
inspections, hesitancy in exploring open spaces that are unfami-
liar and potentially dangerous (i.e. high thigmotaxis), and erratic J. R. Soc. Interface 16: 20190359
swimming patterns dominated by high speed variance
are typically associated with risk aversion and fearful states in
animals [43], including mosquitofish [28,29,36,52,59].
2.3. Experimental procedure
Once a week over seven consecutive weeks, fish were anaesthe-
tized in a solution of tricaine methanesulfonate (MS-222;
168 mg per 1 l H
O), sexed, and their body length (to the nearest
0.5 mm) and body weight (to the nearest 0.01 g) were measured.
These measurements were conducted before the experiment
started (baseline body measurements) and after the conclusion
of each behavioural trial (week 1 to week 6). Fultons condition
factor K(weight length
) [60] was then calcu-
lated as an index for the nutritional state (i.e. body condition)
of each fish at each week.
In each trial, a mosquitofish was gently hand-netted and
placed into an opaque cylinder in the experimental arena for
5 min to allow acclimatization to the set-up. During acclimatiz-
ation, the motors of the robotic platform were turned off and
fish had no visual contact with the apparatus outside the
opaque cylinder. Then, the opaque cylinder was gently removed
and the platform turned on, allowing the fish to explore the arena
in either absence (NP) or presence of the biologically inspired
predator replica (PM, OL1, OL2, CL1 and CL2 conditions) for
15 min. After the trial was completed, the fish was transferred
back into its individual housing cylinder and the next trial was
The behaviour of each individual (n= 75) was tested once a
week over six consecutive weeks, with individuals tested once
per condition. An equal number of replicates were conducted
for each condition. One week interval between two consecutive
behavioural measurements is commonly adopted when testing
individual variation in mosquitofish behaviour to minimize
memory effects [29,52]. Experimental conditions were per-
formed in a randomized order, but the NP condition was
always performed last to mitigate bias on fish baseline behav-
iour caused by individuals being exposed to diverse degrees
of predator threat, as observed in [42] for risk avoidance in mos-
quitofish. Fish were tested in a randomized order to exclude
consistent differences in their behavioural outcome caused by
hunger [61].
2.4. Statistical analysis
We initially tested whether body length, body mass and Fultons
Kwere correlated by estimating phenotypic correlations (i.e.
the overall correlation attributable to between- and within-
individual correlations) with bivariate linear mixed-effects
models (LMMs), as suggested by Dingemanse & Dochtermann
[62]. In these models, we specified the individual as the
random effect (i.e. random intercepts) to account for repeated
measures of the same individual across weeks. Body size was
correlated with both mass and K, while mass and Kwere not
correlated with each other (table S1). Therefore, we included
both body mass and FultonsKas fixed effects in the LMMs
below, while body size was excluded from the models.
Since we were interested in testing whether mosquitofish
antipredator response increased with an increased degree of
biomimicry of the replica, we measured individual behaviour
repeatedly across experimental conditions. We ran separate
LMMs in which distance moved, freezing, speed variance,
mean distance from the replica, predator inspection and thigmo-
taxis were included one-by-one as the dependent variables. In
each model, individual identities were included as the random
effect, while body mass, FultonsK, sex, week and condition
(i.e. the degree of biomimicry of the robotic predator) were
entered as fixed effects. A significant effect of condition in a
given model (or any other fixed effect included in that model)
would indicate that condition explained a significant portion
of the behavioural variance observed after accounting for the
variation explained by the other fixed effects. The significance
of individual differences was tested using both likelihood
ratio tests (LRTs) and Akaike information criteria (AICs), where
a full model including individual as a random effect was
compared with a reduced model in which the random effect
was excluded. Random intercepts represented the proportion
of the total phenotypic variance not attributable to fixed
effects that was explained by among-individual variance, i.e.
differences in personality traits among individuals.
Building upon our previous work [15], we implemented the
information-theoretic notion of transfer entropy to quantify
the influence of the biologically inspired predator replica on
the behaviour of the live fish and vice versa. Given two stochastic
processes, transfer entropy quantifies the reduction in the uncer-
tainty in the prediction of the future of one of the processes
from its present due to additional knowledge about the other
stochastic processes [63]. In this vein, a non-zero value of transfer
entropy indicates a potential influence between the two
processes [63]. Here, transfer entropy was computed on the
time series of the speed of the replica and the mosquitofish,
which were first down-sampled to 1 Hz to ensure that one
time step (1 s) would suffice to encode the response time of the
fish to the replica and vice versa. Therefore, a total of 904
points (904 s) were used per each trial. Then, we converted the
time series into symbols depending on whether the speed
increased or decreased between two consecutive time steps
[64]. In agreement with [15], we computed the transfer entropy
from the replica (R) to the fish (F) as follows:
TErobot!fish ¼X
where Fand Rare the down-sampled time series of the speeds,
and Pr represents the probability mass function computed via
plug-in estimation. By flipping Fwith R, we computed transfer
entropy from the fish to the replica TEfish!robot. Across the five
experimental conditions in which the predator replica was
employed, transfer entropy could only be used in OL1, CL1
and CL2, since the speed of the replica was constant in PM
and OL2 and, therefore, encoded no meaningful information.
We expected information flow in OL1 to be one-directional,
since the replica swam irrespective of the fish, which should
be influenced by the swimming pattern of the replica. On the
other hand, the information flow in CL1 and CL2 was expected
to be two-directional, with the fish responding to the replica and
the replica adjusting its attacks as a function of the behaviour of
the fish. For each of the three conditions (OL1, CL1 and CL2), we
obtained surrogate data from all the possible shuffling (74 × 74)
of the identities of the fish and the replica within each condition.
For each of these shuffling, we randomly selected 74 values
without repetitions to obtain a mean transfer entropy value;
this process was repeated 20000 times to obtain a surrogate
distribution. To ascertain an influence through transfer entropy,
we tested whether the corresponding experimental value was
in the right tail of the distribution. This process was conducted
six times, twice for each of the three conditions to examine
information flow in either direction (fish to robot or robot to
fish, similar to [65]).
We then tested whether transfer entropy differed across con-
ditions and between directions (TEfish!robot and TErobot!fish).
Therefore, we built an LMM with transfer entropy as the depen-
dent variable, the direction of information flow, condition, and
their interaction as fixed effects, and both individual identities
(fish identity) and pair identities ( fish and replica identities)
included as random intercepts. As for the LMMs on behavioural
traits described above, the significance of random intercepts J. R. Soc. Interface 16: 20190359
(both individual and pair identities) was tested using LRTs
and AICs.
Lastly, we were interested in testing whether fish energy
reserves (FultonsK) varied in response to the exposure to
robotic predator replicas. Towards this aim, we built an LMM
with FultonsKas the dependent variable, including individuals
identity as the random effect (i.e. random intercepts) and sex,
week and condition (i.e. the degree of biomimicry of the robotic
predator) as fixed effects. We then tested whether the behaviour-
al variation observed across conditions reflected variation in
FultonsK. Based on our initial hypothesis and findings from
behavioural analyses, experimental conditions were consolidated
in three categories: controls (Kmeasured before the experiment
started, after tests performed in the absence of the predator
replica, and after the exposure to the predator motionless;
baseline, NP and PM, respectively), low degree of biomimicry
(OL1, OL2 and CL2), and high degree of biomimicry (CL1).
Variation in Kwas then tested with an LMM, in which Fultons
Kwas the dependent variable, individualsidentities the random
effect (i.e. random intercepts), and sex, weeks and condition
category the fixed effects.
Data analyses were performed in R v.3.5.1 [66] using the
lme4,nlme,lmerTestand MCMCglmmpackages [6770],
estimated marginal means (EMMs) based on univariate models
and post hoc comparisons were performed with emmeans
adjusted for simultaneous inference with the mvt method [71],
while permutation tests for transfer entropy analysis were
conducted in Matlab (R2018a; MathWorks, Natick, MA, USA
[72]). Prior to all analyses, speed variance was log-transformed
to normalize error distribution in the models residuals. Except
for the permutation test that is independent by error distri-
butions, we assumed Gaussian error distributions that were
confirmed for all response variables after visual inspection of
model residuals. The significance level was set at α< 0.05.
3. Results
Behaviour was strongly dependent on the experimental
condition in which mosquitofish were tested after controlling
for variation explained by week (see results from the LMMs
in table 1). The distance between the fish and the replica
decreased when the replica was allowed to swim in the
arena with respect to the condition PM where it was held
in place ( p< 0.001 in pairwise comparisons between PM
and any other experimental condition; electronic supplemen-
tary material, figure S1). This was especially evident when
attacks were performed in real time by accelerating towards
the fish and the interactive nature of the replica buffered
fish attempts to be away from it ( p< 0.001 in pairwise com-
parisons between CL1 and OL1, OL2 and CL2; electronic
supplementary material, figure S1).
On the contrary, fish tendency to inspect the predator
replica did not vary across swimming replicas, i.e. the
number of inspections in CL1 was undistinguishable from
OL1, OL2 and CL2 (figure 2a). Accordingly, fish swam on
average longer distances, varied their swimming speed
more, and froze less when exposed to a swimming replica
than in control conditions ( p< 0.001 in pairwise comparisons
between NP and PM confronted with any other experimental
condition, except for speed variance and freezing between
NP and CL1 and between PM and CL1, respectively;
electronic supplementary material, figure S1).
Thigmotaxis increased with increasing biomimicry in the
replicas motion, whereby the time interval spent in the
proximity of the walls was longer when fish were exposed
to a replica varying its attacking speed in real time (CL1)
than other replicas ( p< 0.001 in pairwise comparisons
between CL1 and any other experimental condition in
which a robotic replica was employed), with the shortest
time observed in the presence of the motionless replica
(PM; figure 2b). On the other hand, behavioural responses
of fish exposed to an attacking replica that swam at a constant
speed (CL2) were comparable with those observed in OL
conditions (OL1 and OL2), consistently across all measured
traits (figure 2; electronic supplementary material, figure S1).
The variation in body condition (FultonsK) among
individuals was a significant predictor for the variation in
their behavioural response across conditions (see results
from the LMMs in table 1). In particular, individuals with
more energy reserves varied their swimming speed more
(i.e. exhibited higher speed variance) in response to the
replica and an analogous role of Kwas also noted, albeit
not significant, with respect to distance moved, distance
from the replica, predator inspection and thigmotaxis
(table 1). Accordingly, individuals with higher Ktended to
swim longer distances, maintained larger distances from
the replicas, inspected the replicas less, and spent more
time in the proximity of the walls. Nevertheless, we registered
consistent among-individual variance in all traits after that
behavioural variation explained by the model predictors
was accounted for, i.e. fish differed in personality traits (see
results from the LMMs in table 2), except for the mean
distance from the replica and the individual intercepts for
the transfer entropy.
We failed to identify an information transfer flow in the
OL condition OL1 in both directions (i.e. from the robot to
the fish and vice versa; figure 3a,b). On the contrary, a signifi-
cant information transfer was observed in both directions
in CL1 (figure 3c,d) and CL2 ( figure 3e,f). When comparing
information transfers within conditions, we observed that
transfer entropy from the robot to the fish in the OL condition
OL1 was higher than from the fish to the robot ( p= 0.003;
figure 3g), in agreement with our expectations on the one-
directional nature of the interaction in OL1. Transfer entropy
in the closed-loop condition CL1 was also larger from the
robot to the fish than in the opposite direction (p< 0.001),
while transfer entropy in CL2 was comparable between
directions (figure 3g). Importantly, the effect of the replica on
fish behaviour was stronger in CL1 than in CL2 (p=0.042;
figure 3g), while other pairwise comparisons were not signifi-
cant. In other words, the biologically inspired robotic predator
interacting with mosquitofish in real time and accelerating
towards the fish (CL1) was more effective in eliciting anti-
predator responses in mosquitofish than when it attacked at
a constant speed (CL2).
We also found that body condition (FultonsK) varied
across experimental conditions (see results from the LMM
in electronic supplementary material, table S2), with Ksig-
nificantly lower after fish faced the predator replicas than
after fish were tested in the absence of the replica ( p< 0.001
in pairwise comparisons between NP and any other
experimental condition; electronic supplementary material,
figure S2). The decrease in Kafter exposure to the replica
(p< 0.001 in pairwise comparisons between controls and
replicas with either low and high biomimicry; figure 4)
appeared, however, to be independent of the degree of bio-
mimicry of the replica (non-significant pairwise comparison
between low and high biomimicry; figure 4). J. R. Soc. Interface 16: 20190359
4. Discussion
Here, we have disentangled the relative contributions of
swimming pattern and closed-loop control of an interactive
robotic predator on the antipredator behavioural response
and life-history strategies in mosquitofish. Fish thigmotaxis
increased with the degree of biomimicry in the motion of
the replica, suggesting that integrating real-time feedback
Table 1. Analysis of variance with Satterthwaites method from linear mixed models with distance moved, freezing, speed variance, mean distance from replica,
predator inspection, thigmotaxis and transfer entropy as dependent variables. FultonsK, body mass, sex, week and condition are included in all models as xed
factors, except for transfer entropy in which condition, direction, and their interaction were included as xed factors. The individual is included as a random effect
(i.e. random intercepts) in all models, while pair ( shrobot) is included as a second random effect in the transfer entropy model, to account for repeated
measures. The signicance was set at α< 0.05, and signicant results are indicated with * (less than 0.05), ** (less than 0.01) and *** (less than 0.001).
xed factors mean square d.f. Fp-value
distance moved (cm)
K2 801 434 1, 424 3.567 0.059
mass 239 961 1, 121 0.3055 0.581
sex 2 306 162 1, 78 2.936 0.091
week 82 525 615 1, 379 105.080 <0.001***
condition 32 234 416 5, 367 41.044 <0.001***
freezing (s)
K11 392 1, 340 0.359 0.549
mass 57 781 1, 96 1.821 0.180
sex 85 910 1, 75 2.708 0.104
week 3 161 017 1, 386 99.631 <0.001***
condition 735 543 5, 367 23.183 <0.001***
speed variance (cm
K3.376 1, 234 7.044 0.008**
mass 0.047 1, 85 0.099 0.754
sex 0.027 1, 74 0.056 0.814
week 3.881 1, 389 8.099 0.005**
condition 8.935 5, 368 18.644 <0.001***
distance from replica (cm)
K62.46 1, 186 3.784 0.053
mass 0.390 1, 77 0.024 0.878
sex 0.06 1, 73 0.003 0.954
week 0.59 1, 315 0.036 0.850
condition 343.970 4, 291 20.837 <0.001***
predator inspection (counts)
K199.50 1, 288 3.025 0.083
mass 66.50 1, 80 1.008 0.318
sex 212.90 1, 74 3.227 0.076
week 441.80 1, 314 6.698 0.010*
condition 5539.30 4, 291 83.977 <0.001***
thigmotaxis (s)
K59 307 1, 404 2.944 0.087
mass 23 814 1, 107 1.182 0.279
sex 104 813 1, 75 5.202 0.025*
week 232 953 1, 380 11.562 <0.001***
condition 291 277 5, 365 14.457 <0.001***
transfer entropy (bits)
condition <0.001 2, 146 0.514 0.599
direction <0.001 1, 219 49.516 <0.001***
condition × direction <0.001 2, 219 5.262 0.006** J. R. Soc. Interface 16: 20190359
from mosquitofish position in the control of a replica interact-
ing at increasing speed plays a key role in eliciting
antipredator response in mosquitofish. The quantification
of the information flow between the replica and fish sup-
ported the existence of a causal relationship between fish
antipredator response and the motion of the biologically
inspired replica. We also observed that individual behaviour
was relatively predictable, with variations in energy reserves
explaining a large portion of the behavioural variance
observed among mosquitofish. Notably, energy reserves
decreased after fish were exposed to the biologically inspired
robot only 15 min per week, but variation in energy reserves
did not depend on the degree of biomimicry in the motion of
the replica.
After the initial detection of a potential predator, a fish
typically identifies and assesses the threat based on cues
from its natural predators [73]. The extent of an antipredator
response is determined from the trade-off between minimiz-
ing risk of predation and energy consumption towards
survival and reproduction [74], such that greater threats pro-
duce stronger avoidance [75]. Here, we provide experimental
evidence that swimming patterns represent a salient source
of information for predator recognition in mosquitofish
that regulate the extent of their antipredator response. This
evidence is based on highly controllable experiments that
employ a state-of-the-art robotic predator replica, whose
visual appearance and swimming pattern were inspired
by measurements on juvenile largemouth bass, the main
b,c a,b
EMMs (thigmotaxis, s)
EMMs (predator inspection 10 Hz)
Figure 2. Estimated marginal mean (EMM) differences represent adjusted mean differences (+s.e.) in predator inspection (a) and thigmotaxis (b) across conditions
once the contribution of fixed effects included in the model (i.e. FultonsK, body mass, sex, week) is accounted for, except sex that was excluded in EMMs for
predator inspection to preserve positive values in PM condition and favour the interpretation while not altering results. White histograms correspond to control
conditions (NP and PM), light grey histograms to open-loop conditions (OL1 and OL2), and dark grey histograms to closed-loop conditions (CL1 and CL2).NP
condition is not shown for predator inspection (a) since fish were tested in the absence of the predator replica. Means not sharing a common superscript are
significantly different. The significance was set at α< 0.05.
Table 2. Results from general linear mixed models with distance moved, freezing, speed variance, mean distance from replica, predator inspection, thigmotaxis
and transfer entropy as dependent variables. FultonsK, body mass, sex, week and condition are included in all models as xed factors, except for transfer
entropy in which condition, direction, and their interaction were included as xed factors. Random intercepts are included for each individual in all models,
while random intercepts for each pair (shrobot) are included for transfer entropy only, which allowed variance decomposition. Within-individual variance
), among-individual variance (V
) and repeatability are shown for each behavioural trait. Test statistics (
1) and signicant levels of the random
effects (i.e. intercepts) were estimated using LRTs (p) and AICs between the full and the null model. Note that ΔAIC corresponds to the difference in AIC
between the null models minus the AIC from the full model. The signicance was set at α< 0.05, and signicant results are indicated with * (less than 0.05)
and *** (less than 0.001).
variance components V
repeatability ΔAIC
distance moved (cm) 785 359 599 257 0.433 118.770 120.769 <0.001***
freezing (s) 31 727 9038 0.222 33.978 35.977 <0.001***
speed variance (cm
) 0.479 0.035 0.069 2.322 4.322 0.038*
distance from replica (cm) 16.507 0.482 0.028 1.461 0.539 0.463
predator inspection (counts) 65.960 19.270 0.226 25.019 27.019 <0.001***
thigmotaxis (s) 20 148 11 185 0.357 80.800 82.800 <0.001***
transfer entropyindividual (bits) <0.001 <0.001 0.172 1.7 3.734 0.053
transfer entropypair (bits) <0.001 <0.001 0.455 42.100 44.068 <0.001*** J. R. Soc. Interface 16: 20190359
predator of mosquitofish in the wild [3941]. Not only did the
robotic replica allow for controlling the swimming speed and
acceleration of the predator stimulus, but also it afforded the
implementation of controlled attacks towards mosquitofish to
study their antipredator response in real time. By opting for a
robotics-based platform in lieu of a live predator, we were
able to exclude potential correlations between antipredator
response of mosquitofish and inherent biological variations
in the predator behaviour (i.e. idiosyncrasies with focal
individuals, fatigue and hunger) that could confound
hypothesis testing.
The more robust antipredator behavioural response was
registered when mosquitofish were exposed to a replica
swimming at a varying speed and performing targeted, fast
attacks. Reducing the degree of biomimicry towards a replica
that performed attacks in real time at a constant speed
resulted into a weaker antipredator behavioural response,
similar to that registered with non-interactive replicas that
followed predetermined swimming trajectories. This evi-
dence aligns with prediction from the literature positing
that speed and acceleration should play a key role on prey
predator interactions in fish [76]. Our information-theoretic
analysis of the interaction between the robotic replica and
the fish suggests the presence of a causeeffect relationship
underlying the antipredator behavioural response of mosquito-
fish, which confirms the expected link between a predators
attacking speed and fish behavioural response [77]. More
specifically, we determined that the uncertainty in the predic-
tion of the future speed of mosquitofish from its present
speed was reduced due to additional knowledge about the
0 1.5 3.0 4.5 6.0
×10–3 0 1.5 3.0 4.5 6.0 ×10–3
TE (bits)
probability probabilityprobability
surrogate data
real TE
95% quantile
Figure 3. Transfer entropy between fish and robotic replicas. Transfer entropy from fish to robot is represented in (a,c,e) and from robot to fish in (b,d,f) with
respect to conditions OL1 (first row), CL2 (second row) and CL1 (third row). Transfer entropy in both directions (fish-to-robot and robot-to-fish) for each of the three
conditions is represented in (g) (+s.e.). Means not sharing a common superscript are significantly different. The significance was set at α< 0.05. (Online version in
colour.) J. R. Soc. Interface 16: 20190359
speed of the replica, such that the motion of the replica encoded
valuable information about the behaviour of mosquitofish.
Beyond the analysis of the mean behavioural response at
the population level, we discovered that a relatively short
exposure to the biologically inspired robotic predator (only
15 min per week) resulted in a substantial reduction in the
whole body condition of mosquitofish (index of fat reserves
for a given body size; condition factor K) that did not
depend on the swimming pattern of the robot. Recent evi-
dence from multiple populations of mosquitofish in the
wild has shown that the condition factor Kin mosquitofish
decreased on average 5.8% over a five month period in
response to severe environmental challenges associated with
water pollution [78]. Here, we observed that the body con-
dition declined 3.1% over a week, after mosquitofish were
exposed once to a predator replica, thereby suggesting a
hidden effect of the robot on mosquitofish life-history adjust-
ments. This finding aligns with evidence of nonlethal effects
of predatorprey interactions [26], whereby costs of antipreda-
tor responses extend to ecologically relevant traits beyond
behaviour, such as physiology and body condition [79].
In fact, theory predicts that stress responses affect the way
animals allocate resources to fuel emergency functions [79],
with animals investing relatively more energy in survival
(i.e. escaping from the predator) and relatively less in future
reproduction (i.e. energy reserves) with increasing predation
risk [80]. With respect to mosquitofish, nonlethal effects of
predator exposure have been found to lower their body con-
dition, ultimately leading to lower fertility and fecundity
rates [31]. Under this perspective, evidence from this study
indicates that a relatively brief exposure to a biologically
inspired robotic predator compromised the body condition
of mosquitofish. Notably, the body condition increased
again when mosquitofish were tested in the arena in the
absence of the replica, indicating that variation in body con-
dition resulted from the exposure to the robotic predator
rather than other factors (for example, time, exposure to the
arena and handling of the fish).
At the individual level, we found that fish differed consist-
ently from each other in the extent of their antipredator
response across six repeated exposures to robotic predators
varying in their degree of biomimicry (i.e. fish differed in
personality traits [43]). While the presence of personality vari-
ation among mosquitofish is well documented in the literature
(see for example [29,52,59] and references therein), this study
offers evidence that meaningful variation in antipredator
response among mosquitofish can be successfully captured
using robotic stimuli. Interestingly, a large portion of the
variance in the antipredator response observed among mos-
quitofish was explained by variation in their body condition.
In particular, individuals in better body conditions varied
their swimming speed more in response to the robotic preda-
tor, tended to swim longer distances, maintained larger
distances and inspected less the replicas, and spent more
time in the proximity of the wall than mosquitofish in poorer
body conditions. Individuals can trade-off survival at the
cost of future reproduction, but the antipredator behavioural
response of an individual should also depend on its body con-
dition [81] as the reproductive value is condition-dependent.
In this vein, our results are in agreement with predictions
from the life-history theory that individuals with high future
expectations (i.e. individuals with high energy reserves)
should systematically be more risk-averse than others [81].
Therefore, our findings suggest that antipredator behavioural
response towards robotic predator fish differs at the individual
level in a relatively predictable manner.
This study contributes to the state of the art on the modu-
lation of the behaviour of invasive and pest species through
the use of predator-mimicking robotic fish [19,28], supporting
the technological evolution of pest control agents, along
similar line of development as insects [50] and birds [51].
Specifically, we aimed at the precise quantification of granular
features of predator locomotion on antipredator responses of
invasive mosquitofish through the development of a state-
of-the-art robotic predator whose swimming characteristics
can be controlled across a continuum range of biomimicry.
Our findings build on previous research efforts on the modu-
lation of mosquitofish behaviour through biologically inspired
robots, shedding light on the role of the robot morphology on
mosquitofish behaviour [28] and addressing the differential
response of mosquitofish and zebrafish to robots [19]. In par-
ticular, we demonstrated that a biologically inspired robotic
predator swimming at a varying speed and performing
targeted attacks elicits a strong antipredator behavioural
response that erodes energy reserves and compromises the
body condition of mosquitofish. We propose that further
efforts should test whether biologically inspired robots can
effectively represent a novel, autonomous, and effective sol-
ution to contrast the negative impact of invasive
mosquitofish on freshwater ecosystems worldwide [59].
Ethics. Experiments were performed in accordance with relevant
guidelines and regulations and were approved by the University
Animal Welfare Committee (UAWC) of New York University
under the protocol number 13-1424. Notably, pilot tests on predator
fish were approved through an animal care permit (G 0074/15)
granted by the Landesamt für Gesundheit und Soziales Berlin
(LAGeSo) and performed in Germany.
Data accessibility. All data can be found at
EMMs (K, g mm–3 104)
1.8 abb
low biomimicry high biomimicry
Figure 4. Estimated marginal mean (EMM) differences represent adjusted
mean differences (+s.e.) in Fultons condition factor Kacross conditions
once the contribution of fixed effects included in the model (i.e. sex and
week) is accounted for. The white histogram corresponds to controls (base-
line, NP and PM), the light grey histogram to replicas with low biomimicry
(OL1, OL2 and CL2), and the dark grey histogram to replicas with high bio-
mimicry (CL1). Means not sharing a common superscript are significantly
different. The significance was set at α< 0.05. J. R. Soc. Interface 16: 20190359
Authorscontributions. G.P. and M.P. conceived the research question and
supervised the research. G.P. designed the experiment. G.P., V.R.S.
and C.S. developed the experimental set-up. M.K., V.R.S. and C.S.
conducted the experiments. G.P. and M.K. analysed the data and
all authors discussed the results. G.P. and M.K. wrote a first draft
of the Material and Methods section. G.P. and M.P. wrote the
manuscript. All the authors reviewed the final draft.
Competing interests. We declare we have no competing interests.
Funding. This work was supported by the National Science Foundation
under grant nos. CMMI-1433670 and CMMI-1505832 and by the
Forrest Research Foundation.
Acknowledgements. The authors are grateful to Yanpeng Yang for his
support with the setting of the robotic platform, Shinnosuke
Nakayama for advising on the statistics, and the anonymous
reviewers for their constructive feedback that has helped improve
the work and the presentation.
1. Pimentel D, Zuniga R, Morrison D. 2005 Update on
the environmental and economic costs associated
with alien-invasive species in the United States.
Ecol. Econ. 52, 273288. (doi:10.1016/j.ecolecon.
2. Raven PH, Johnson GB. 1992 Biology, 3rd edn.
St Louis, MO: Mosby Year Book.
3. Byers JE et al. 2002 Directing research to
reduce the impacts of nonindigenous species.
Conserv. Biol. 16, 630640. (doi:10.1046/j.1523-
4. Sala OE et al. 2000 Global biodiversity scenarios for
the year 2100. Science 287, 17701774. (doi:10.
5. Pyke GH. 2005 A review of the biology of Gambusia
affinis and G. holbrooki.Rev. Fish Biol. Fisheries 15,
339365. (doi:10.1007/s11160-006-6394-x)
6. Pyke G. 2008 Plague minnow or mosquito fish?
A review of the biology and impacts of introduced
Gambusia species. Annu. Rev. Ecol. Evol. Syst. 39,
171191. (doi:10.1146/annurev.ecolsys.39.110707.
7. Leprieur F, Beauchard O, Blanchet S, Oberdorff T,
Brosse S. 2008 Fish invasions in the worlds river
systems: when natural processes are blurred by
human activities. PLoS Biol. 6, e28. (doi:10.1371/
8. Kats LB, Ferrer RP. 2003 Alien predators and
amphibian declines: review of two decades of science
and the transition to conservation. Divers. Distributions
9,99110. (doi:10.1046/j.1472-4642.2003.00013.x)
9. Lowe S, Browne M, Boudjelas S, De Poorter M. 2000
100 of the worlds worst invasive alien species: a
selection from the global invasive species database,
vol. 12. Auckland, New Zealand: Invasive Species
Specialist Group.
10. Westhoff JT, Watts AV, Mattingly HT. 2013 Efficacy of
artificial refuge to enhance survival of young
Barrens topminnows exposed to western
mosquitofish. Aquat. Conserv. Marine Freshwater
Ecosyst. 23,6576. (doi:10.1002/aqc.2265)
11. Willis K, Ling N. 2000 Sensitivities of mosquitofish
and black mudfish to a piscicide: could rotenone be
used to control mosquitofish in New Zealand
wetlands? N.Z. J. Zool. 27,8591. (doi:10.1080/
12. Brookhouse N, Coughran J. 2010 Exploring the
potential for an ecology-specific, physical control
method of the exotic pest mosquitofish, Gambusia
holbrooki.Ecol. Manage. Restor. 11, 226228.
13. Romano D, Donati E, Benelli G, Stefanini C. 2018 A
review on animalrobot interaction: from bio-hybrid
organisms to mixed societies. Biol. Cybern. 113,
201225. (doi:10.1007/s00422-018-0787-5)
14. Krause J, Winfield AFT, Deneubourg JL. 2011
Interactive robots in experimental biology. Trends Ecol.
Evol. 26,369375. (doi:10.1016/j.tree.2011.03.015)
15. Porfiri M. 2018 Inferring causal relationships in
zebrafish-robot interactions through transfer
entropy: a small lure to catch a big fish. Anim.
Behav. Cogn. 5, 341367. (doi:10.26451/abc.05.04.
16. Romano D, Benelli G, Donati E, Remorini D, Canale
A, Stefanini C. 2017 Multiple cues produced by a
robotic fish modulate aggressive behaviour in
Siamese fighting fishes. Sci. Rep. 7, 4667. (doi:10.
17. Romano D, Benelli G, Hwang JS, Stefanini C. 2019
Fighting fish love robots: mate discrimination in
males of a highly territorial fish by using female-
mimicking robotic cues. Hydrobiologia 833,
185196. (doi:10.1007/s10750-019-3899-6)
18. Polverino G, Phamduy P, Porfiri M. 2013 Fish and
robots swimming together in a water tunnel: robot
color and tail-beat frequency influence fish
behavior. PLoS ONE 8, e77589. (doi:10.1371/journal.
19. Polverino G, Porfiri M. 2013 Zebrafish (Danio rerio)
behavioural response to bioinspired robotic fish and
mosquitofish (Gambusia affinis). Bioinspir. Biomim.
8, 044001. (doi:10.1088/1748-3182/8/4/044001)
20. Phamduy P, Polverino G, Fuller RC, Porfiri M. 2014
Fish and robot dancing together: bluefin killifish
females respond differently to the courtship of a
robot with varying color morphs. Bioinspir. Biomim.
9, 036021. (doi:10.1088/1748-3182/9/3/036021)
21. Marras S, Porfiri M. 2012 Fish and robots swimming
together: attraction towards the robot demands
biomimetic locomotion. J. R. Soc. Interface 9,
18561868. (doi:10.1098/rsif.2012.0084)
22. Neri D, Ruberto T, Cord-Cruz G, Porfiri M. 2017
Information theory and robotics meet to study
predator-prey interactions. Chaos: Interdiscip.
J. Nonlinear Sci. 27, 073111. (doi:10.1063/1.4990051)
23. Ladu F, Bartolini T, Panitz SG, Chiarotti F, Butail S,
Macrì S, Porfiri M. 2015 Live predators, robots, and
computer-animated images elicit differential
avoidance responses in zebrafish. Zebrafish 12,
205214. (doi:10.1089/zeb.2014.1041)
24. Porfiri M, Spinello C, Yang Y, Macrì S. 2019 Zebrafish
adjust their behavior in response to an interactive
robotic predator. Front. Robot. AI 6, 38. (doi:10.
25. Swain DT, Couzin ID, Leonard NE. 2011 Real-time
feedback-controlled robotic fish for behavioral
experiments with fish schools. Proc. IEEE 100,
150163. (doi:10.1109/JPROC.2011.2165449)
26. Lima SL. 1998 Nonlethal effects in the ecology of
predator-prey interactions. Bioscience 48,2534.
27. Lima SL, Dill LM. 1990 Behavioral decisions made
under the risk of predation: a review and prospectus.
Can. J. Zool. 68,619640. (doi:10.1139/z90-092)
28. Polverino G, Porfiri M. 2013 Mosquitofish (Gambusia
affinis) responds differentially to a robotic fish of
varying swimming depth and aspectratio. Behav.Brain
Res. 250,133138. (doi:10.1016/j.bbr.2013.05.008)
29. Polverino G, Santostefano F, Díaz-Gil C, Mehner T.
2018 Ecological conditions drive pace-of-life
syndromes by shaping relationships between life
history, physiology and behaviour in two
populations of eastern mosquitofish. Sci. Rep. 8,
14673. (doi:10.1038/s41598-018-33047-0)
30. Heinen-Kay JL, Schmidt DA, Stafford AT, Costa MT,
Peterson MN, Kern EM, Langerhans RB. 2016
Predicting multifarious behavioural divergence in
the wild. Anim. Behav. 121,310. (doi:10.1016/j.
31. Mukherjee S, Heithaus MR, Trexler JC, Ray-Mukherjee
J, Vaudo J. 2014 Perceived risk of predation affects
reproductive life-history traits in Gambusia holbrooki,
but not in Heterandria formos.PLoS ONE 9, e88832.
32. Benejam L, Alcaraz C, Sasal P, Simon-Levert G,
García-Berthou E. 2009 Life history and parasites of
the invasive mosquitofish (Gambusia holbrooki)
along a latitudinal gradient. Biol. Invasions 11,
22652277. (doi:10.1007/s10530-008-9413-0)
33. Langerhans RB, Layman CA, Shokrollahi A, DeWitt
TJ. 2004 Predator-driven phenotypic diversification
in Gambusia affinis.Evolution 58, 23052318.
34. Hartman EJ, Abrahams MV. 2000 Sensory compensation
and the detection of predators: the interaction between
chemical and visual information. Proc. R. Soc. Lond. B
267,571575. (doi:10.1098/rspb.2000.1039)
35. Ward AJW, Mehner T. 2010 Multimodal mixed
messages: the use of multiple cues allows greater
accuracy in social recognition and predator detection
decisions in the mosquitofish, Gambusia holbrooki.
Behav. Ecol. 21, 13151320. (doi:10.1093/beheco/
arq152) J. R. Soc. Interface 16: 20190359
36. Polverino G, Liao JC, Porfiri M. 2013 Mosquitofish
(Gambusia affinis) preference and behavioral response
to animated images of conspecifics altered in their
color, aspect ratio, and swimming depth. PLoS ONE 8,
e54315. (doi:10.1371/journal.pone.0054315)
37. Casner AM, Fackelman HC, Degtyareva O, Kight SL.
2016 Do female western mosquitofish, Gambusia
affinis, prefer ornaments that males lack? Ethology
122, 561570. (doi:10.1111/eth.12507)
38. Chen BJ, Liu K, Zhou LJ, Gomes-Silva G, Sommer-
Trembo C, Plath M. 2018 Personality differentially affects
individual mate choice decisions in female and male
western mosquitofish (Gambusia affinis). PLoS ONE 13,
e0197197. (doi:10.1371/journal.pone.0197197)
39. García-Berthou E. 2002 Ontogenetic diet shifts and
interrupted piscivory in introduced largemouth bass
(Micropterus salmoides). Int. Rev. Hydrobiol. 87,
353363. (doi:10.1002/1522-2632(200207)87:4
40. Godinho FN, Ferreira MT, Cortes RV. 1997 The
environmental basis of diet variation in
pumpkinseed sunfish, Lepomis gibbosus, and
largemouth bass, Micropterus salmoides, along an
Iberian river basin. Environ. Biol. Fishes 50,
105115. (doi:10.1023/A:1007302718072)
41. Wicker AM, Johnson WE. 1987 Relationships among
fat content, condition factor, and first-year survival
of Florida largemouth bass. Trans. Am. Fish. Soc.
116, 264271. (doi:10.1577/1548-8659(1987)116<
42. Darby N, McGhee K. 2019 Boldness is affected by
recent experience with predation cues and body size
in mosquitofish. Behav. Processes 164, 143149.
43. Réale D, Reader SM, Sol D, McDougall PT, Dingemanse
NJ. 2007 Integrating animal temperament within
ecology and evolution. Biol. Rev. 82,291318. (doi:10.
44. Wolf M, van Doorn GS, Leimar O, Weissing FJ. 2007
Life-history trade-offs favour the evolution of
animal personalities. Nature 447, 581584. (doi:10.
45. Rossi C, Coral W, Barrientos A. 2012 Robotic fish to
lead the school. In Swimming physiology of fish:
towards using exercise for farming a fit fish in
sustainable aquaculture (eds AP Palstra, JV Planas).
New York, NY: Springer.
46. Schreiber T. 2000 Measuring information transfer. Phys.
Rev. Lett. 85, 461. (doi:10.1103/PhysRevLett.85.461)
47. Ruberto T, Polverino G, Porfiri M. 2017 How
different is a 3D-printed replica from a conspecific in
the eyes of a zebrafish? J. Exp. Anal. Behav. 107,
279293. (doi:10.1002/jeab.247)
48. Bierbach D, Landgraf T, Romanczuk P, Lukas J,
Nguyen H, Wolf M, Krause J. 2018 Using a robotic
fish to investigate individual differences in social
responsiveness in the guppy. R. Soc. open sci. 5,
181026. (doi:10.1098/rsos.181026)
49. Romano D, Benelli G, Stefanini C. 2019 Encoding
lateralization of jump kinematics and eye use in a
locust via bio-robotic artifacts. J. Exp. Biol. 222,
jeb187427. (doi:10.1242/jeb.187427)
50. Romano D, Benelli G, Stefanini C. 2017 Escape and
surveillance asymmetries in locusts exposed to a
Guinea fowl-mimicking robot predator. Sci. Rep. 7,
12825. (doi:10.1038/s41598-017-12941-z)
51. Folkertsma GA, Straatman W, Nijenhuis N, Venner
CH, Stramigioli S. 2017 Robird: a robotic bird of
prey. IEEE Rob. Autom. Mag. 24,2229. (doi:10.
52. Polverino G, Cigliano C, Nakayama S, Mehner T.
2016 Emergence and development of personality
over the ontogeny of fish in absence of
environmental stress factors. Behav. Ecol.
Sociobiol. 70, 20272037. (doi:10.1007/
53. DeLellis P, Cadolini E, Croce A, Yang Y, di Bernardo
M, Porfiri M. Submitted. Model-based feedback
control of live zebrafish behavior via interaction
with a robotic replica.
54. Yang Y, Clément RJG, Ghirlanda S, Porfiri M. 2019 A
comparison of individuallearning and social learning in
zebrafish through an ethorobotics approach. Front.
Robot. AI 6, 71. (doi:10.3389/frobt.2019.00071)
55. Butail S, Bartolini T, Porfiri M. 2013 Collective
response of zebrafish shoals to a free-swimming
robotic fish. PLoS ONE 8, e76123. (doi:10.1371/
56. Kolok AS. 1992 The swimming performances of
individual largemouth bass (Micropterus salmoides)
are repeatable. J. Exp. Biol. 170, 265270.
57. Magurran AE, Seghers BH. 1990 Population
differences in predator recognition and attack cone
avoidance in the guppy Poecilia reticulata.Anim.
Behav. 40, 443452. (doi:10.1016/S0003-
58. Macrì S, Neri D, Ruberto T, Mwaffo V, Butail S,
Porfiri M. 2017 Three-dimensional scoring of
zebrafish behavior unveils biological phenomena
hidden by two-dimensional analyses. Sci. Rep. 7,
1962. (doi:10.1038/s41598-017-01990-z)
59. Polverino G, Ruberto T, Staaks G, Mehner T. 2016
Tank size alters mean behaviours and individual
rank orders in personality traits of fish depending
on their life stage. Anim. Behav. 115, 127135.
60. Froese R. 2006 Cube law, condition factor and
weightlength relationships: history, meta-analysis
and recommendations. J. Appl. Ichthyol. 22,
241253. (doi:10.1111/j.1439-0426.2006.00805.x)
61. Krause J, Loader SP, McDermott J, Ruxton GD. 1998
Refuge use by fish as a function of body length-
related metabolic expenditure and predation risks.
Proc. R. Soc. Lond. B 265, 23732379. (doi:10.1098/
62. Dingemanse NJ, Dochtermann NA. 2013 Quantifying
individual variation in behaviour: mixed-effect
modelling approaches. J. Anim. Ecol. 82,3954.
63. Bossomaier T, Barnett L, Harré M, Lizier JT. 2016 An
introduction to transfer entropy. Cham, Switzerland:
Springer International Publishing.
64. Porfiri M, Ruiz Marín M. 2018 Inference of time-
varying networks through transfer entropy, the case
of a Boolean network model. Chaos:
Interdiscip. J. Nonlinear Sci. 28, 103123. (doi:10.
65. Zhang P, Krasner E, Peterson SD, Porfiri M. 2019 An
information-theoretic study of fish swimming in the
wake of a pitching airfoil. Physica D 396,3546.
66. R Core Team. 2016 R: a language and environment
for statistical computing. Vienna, Austria: R
Foundation for Statistical Computing.
67. Bates D, Maechler M, Bolker B, Walker S. 2014
lme4: Linear mixed-effects models using Eigen and
S4. R package version 1.1-5.
68. Pinheiro J, Bates D, DebRoy S, Sarkar D, CoreTeam
R. 2014 nlme: linear and nonlinear mixed effects
models. R package version 3.1-118.
69. Kuznetsova A, Brockhoff PB, Christensen RHB. 2013
lmerTest: tests for random and fixed effects for
linear mixed effect models (lmer objects of lme4
package). R package version 2-6.
70. Hadfield JD. 2010 MCMC methods for multi-
response generalized linear mixed models: the
MCMCglmm R package. J. Stat. Softw. 33,122.
71. Lenth R. 2018 Emmeans: estimated marginal means,
aka least-squares means. R package version 1.
72. Nichols TE, Holmes AP. 2002 Nonparametric
permutation tests for functional neuroimaging: a
primer with examples. Hum. Brain Mapp. 15,125.
73. Kelly JL, Magurran AE. 2003 Learned predator
recognition and antipredator responses in fishes.
Fish Fisheries 4, 216226. (doi:10.1046/j.1467-
74. Ydenberg RC, Dill LM. 1986 The economics of
fleeing from predators. Adv. Study Behav. 16,
229249. (doi:10.1016/S0065-3454(08)60192-8)
75. Helfman GS. 1989 Threat-sensitive predator
avoidance in damselfish-trumpetfish interactions.
Behav. Ecol. Sociobiol. 24,4758. (doi:10.1007/
76. Domenici P. 2001 The scaling of locomotor
performance in predator-prey encounters: from fish
to killer whales. Comp. Biochem. Physiol. A Mole.
Integr. Physiol. 131, 169182. (doi:10.1016/S1095-
77. Meager JJ, Domenici P, Shingles A, Utne-Palm AC.
2006 Escape responses in juvenile Atlantic cod
Gadus morhua L.: the effects of turbidity and
predator speed. J. Exp. Biol. 209, 41744184.
78. Rautenberg GE, Amé MV, Monferrán MV, Bonansea
RI, Hued AC. 2015 A multi-level approach using
Gambusia affinis as a bioindicator of environmental
pollution in the middle-lower basin of Suquía River.
Ecol. Indic. 48, 706720. (doi:10.1016/j.ecolind.
79. Hawlena D, Schmitz OJ. 2010 Physiological stress as
a fundamental mechanism linking predation to
ecosystem functioning. Am. Nat. 176, 537556.
80. Stearns SC. 1992 The evolution of life histories.
Oxford, UK: Oxford University Press.
81. Clark CW. 1994 Antipredator behavior and the asset-
protection principle. Behav. Ecol. 5, 159170.
(doi:10.1093/beheco/5.2.159) J. R. Soc. Interface 16: 20190359
... In particular, robotics is emerging as a promising approach to studying animal behavior (Krause et al., 2011;Romano et al., 2018;Webb, 2000) and animal invasions (Dufour et al., 2020;Polverino et al., 2019). State-of-the-art robots can move autonomously (Butail et al., 2013;Butail et al., 2014;Gribovskiy et al., 2018;Katzschmann et al., 2018), simulate specific characteristics of live fish (Landgraf et al., 2016;Phamduy et al., 2014;Polverino et al., 2012;Romano et al., 2017a), infiltrate social groups (Bierbach et al., 2018;Butail et al., 2013;Marras and Porfiri, 2012), and interact with live animals in real time (Bonnet et al., 2019;Kopman et al., 2013;Polverino et al., 2019a). Recent studies underscore the extraordinary opportunity of using robots inspired by live predators to control the behavior of pests such as locusts (Romano et al., 2017b(Romano et al., , 2019) that pose a threat to human agricultural economies. ...
... Efforts from our group have shown that biologicallyinspired robotic predators can repel invasive mosquitofish (Polverino and Porfiri, 2013a), undermine their health (Polverino et al., 2019a), and simultaneously attract noninvasive fish species (Polverino and Porfiri, 2013b). ...
... The knowledge that altering the behavior of fish has repercussions on their physiology, growth, fertility, and reproduction is central for studying the developmental and evolutionary consequences of behavioral manipulation. Yet whether exposure to biologically-inspired robots has effects that extend beyond the behavior of animals remains unexplored (but see Polverino et al., 2019a). ...
Conference Paper
Robotics is emerging as a promising approach to study animal behavior. Biologically-inspired robots can manipulate the behavior of live animals and are increasingly employed to uncover the underpinnings of sociality and mate choice in the animal realm. But behavioral variation between animals plays a critical role for their ecology and evolution, and ultimately it determines variation in the survival, growth, and reproduction of individuals. While the study of behavioral responses of animals toward their robotic counterparts dominates the literature, it remains largely untested whether the life-history strategies of live animals can be artificially manipulated with biologically-inspired robots. Recently, predator-mimicking robots allowed to successfully study antipredator responses of highly invasive fish in detail, revealing that costs of behavioral alternations induced by robotic predators can impact the health and survival of invaders. The evidence that biologically-inspired robots can undermine the ecological success of invasive animals opens the door to novel experimental analyses at the interface between robotics, ecology, and invasion biology.
... Studying interactions between fish and robots is not a new idea. Key results in studying fish-robot interaction have been achieved using robots that either react to surrounding fish behavior (Kim et al., 2018;Polverino et al., 2019;Cazenille et al., 2018a,b) or operate without closed loop feedback from other agents (Landgraf et al., 2016;Utter and Brown, 2020;Cazenille et al., 2018b;Coulson et al., 2018;Abaid et al., 2013;Romano et al., 2019a)). Only two fish-robot interaction studies have included archerfish (Coulson et al., 2018;Utter and Brown, 2020). ...
... Establishing whether the attractiveness of a robotic agent to archerfish (in a leader-follower sense: see Porfiri (2018)) could help optimize a robot design for use in studies that specifically investigate social effects on archerfish hunting behavior. While robot activity level has been the focus of some prior work in the literature using various fish species including zebrafish (Ruberto et al., 2016;Polverino et al., 2019), a similar study has not been conducted with archerfish, and studies that explicitly vary the parameters of a stochastic robot motion model and map these variations to changes in fish behavior are rare. Our work uses a simplistic but naturalistic motion model for the robot, given that naturalistic, stochastic motion has been a feature of several recent studies of fish-robot interaction (Cazenille et al., 2018a,b). ...
... Recently, the use of biomimetic robots is advancing our knowledge on animal cognition and behavioural ecology [15][16][17][18][19]. This emergent field of biorobotics and bionics is based on the interaction between living organisms and artificial agents, establishing a biohybrid system [20][21][22]. ...
Collective behaviours in homogeneous shoals provide several benefits to conspecifics, although mixed-species aggregations have been reported to often occur. Mixed aggregations may confer several beneficial effects such as antipredator and foraging advantages. However, the mechanisms promoting phenotypically heterogeneous fish aggregations have been poorly explored so far. Herein, the neon tetra Paracheirodon innesi was selected as ideal model organism to test the role of visible phenotypic traits in promoting fish shoaling. Robotic fish replicas of different colour, but with the morphology inspired to P. innesi, were developed to test the affiliation behaviour of neon tetra individuals towards fish replicas with different phenotypic traits. P. innesi individuals showed a decreasing preference in shoaling with the biomimetic, the blue, the red, and the grey replicas. This could be due to the greater visibility of the blue colour even in dark conditions. Furthermore, an increased reddening of the livery is often caused by physiological processes related to a non-optimal behavioural status. The time spent in shoaling with each fish replica was strongly influenced by different ecological contexts. The longest shoaling duration was observed when a biomimetic predator was present, while the shortest shoaling duration was recorded in presence of food. This confirms the hypothesis that heterogeneous shoals are promoted by the anti-predator benefits, and reduced by competition. This study allowed to understand basic features of the behavioural ecology favouring heterogeneous aggregations in shoaling fish, and provided a novel paradigm for biohybrid robotics.
... Robotics applied to ethological studies, obviously, offers capabilities extending far beyond the previously mentioned methods, allowing one to create biomimetic artificial agents performing complex behaviors that require sophisticated locomotion patterns or coordinated moving (Butail et al., 2014;Phamduy et al., 2014;Abdai et al., 2018;Bierbach et al., 2020;Romano and Stefanini, 2021a), and producing several kind of signals (Partan et al., 2009;Polverino et al., 2019;Romano and Stefanini, 2021b). Biomimetic agents can be fully controlled, allowing one to test all subjects with an identical set of cues. ...
Full-text available
In so-called ethorobotics and robot-supported social cognitive neurosciences, robots are used as scientific tools to study animal behavior and cognition. Building on previous epistemological analyses of biorobotics, in this article it is argued that these two research fields, widely differing from one another in the kinds of robots involved and in the research questions addressed, share a common methodology, which significantly differs from the “synthetic method” that, until recently, dominated biorobotics. The methodological novelty of this strategy, the research opportunities that it opens, and the theoretical and technological challenges that it gives rise to, will be discussed with reference to the peculiarities of the two research fields. Some broad methodological issues related to the generalization of results concerning robot-animal interaction to theoretical conclusions on animal-animal interaction will be identified and discussed.
... The use of robotic fish mimicking conspecifics of P. innesi can be useful to understand which intrinsic and extrinsic mechanisms cause anxiety, as well as how social robots can be effectively used as pathological anxiety treatments. Animal-robot interactions and ethorobotics are advanced biorobotic and bionic paradigms merging robotics with ethology that enable to establish biohybrid social systems useful for multidisciplinary purposes [29][30][31][32][33][34][35][36][37]. ...
Full-text available
The emerging field of social robotics comprises several multidisciplinary applications. Anxiety and stress therapies can greatly benefit by socio-emotional support provided by robots, although the intervention of social robots as effective treatment needs to be fully understood. Herein, Paracheirodon innesi , a social fish species, was used to interact with a robotic fish to understand intrinsic and extrinsic mechanisms causing anxiety, and how social robots can be effectively used as anxiety treatments. In the first experiment we tested the effects of a conspecific-mimicking robot on the fish tendency to swim in the bottom when transferred in a new tank. Here, P. innesi spent a significantly longer time in the upper section of the test tank when the robotic fish was present, clearly indicating a reduction of their state of anxiety due to social stimuli. The second experiment was based on a modification of the dark/light preference test, since many teleost fish are scototactic, preferring dark environments. However, when the robotic fish was placed in the white half of the test tank, P. innesi individuals swam longer in this section otherwise aversive. Social support provided by the robotic fish in both experiments produced a better recovery from anxiety due to social buffering, a phenomenon regulated by specific neural mechanisms. This study provides new insights on the evolution and mechanisms of social buffering to reduce anxiety, as well as on the use of social robots as an alternative to traditional approaches in treating anxiety symptoms.
... Robots enable the empirical implementation of some theoretical models of interactions and movements to test hypotheses from fields like movement ecology (collective movement behavior, Faria et al. 2010, Butail et al. 2014, Butail et al. 2016, Bonnet et al. 2018, Jolles et al. 2020, sexual selection (mate choice and aggression, Phamduy et al. 2014, Romano et al. 2017 or natural selection (predatorprey interactions, Swain et al. 2012, Abaid et al. 2013, Heathcote et al. 2020, Polverino et al. 2019. These examples show that especially fishes have been used in experimentation with biomimetic robots and we thus focus on fish to elaborate on the benefits of biomimetic robots for an implementation of the 3Rs. ...
... However, little is known on the mechanisms promoting phenotypically heterogeneous fish aggregations, especially due to the difficulty in controlling different ecological contexts, and in testing separate stimuli. Recently, the use of biomimetic robots is advancing our knowledge on animal cognition and behavioural ecology (Polverino, et al. 2019;Bierbach, et al. 2020;Romano and Stefanini, 2021a). This emergent field of biorobotics and bionics is based on the interaction between living organisms and artificial agents, establishing a biohybrid system Romano, et al. 2019;Bonnet, et al. 2019). ...
Synopsis Comparative biologists have typically used one or more of the following methods to assist in evaluating the proposed functional and performance significance of individual traits: comparative phylogenetic analysis, direct interspecific comparison among species, genetic modification, experimental alteration of morphology (for example by surgically modifying traits), and ecological manipulation where individual organisms are transplanted to a different environment. But comparing organisms as the endpoints of an evolutionary process involves the ceteris paribus assumption: that all traits other than the one(s) of interest are held constant. In a properly controlled experimental study, only the variable of interest changes among the groups being compared. The theme of this paper is that the use of robotic or mechanical models offers an additional tool in comparative biology that helps to minimize the effect of uncontrolled variables by allowing direct manipulation of the trait of interest against a constant background. The structure and movement pattern of mechanical devices can be altered in ways not possible in studies of living animals, facilitating testing hypotheses of the functional and performance significant of individual traits. Robotic models of organismal design are particularly useful in three arenas: (1) controlling variation to allow modification only of the trait of interest, (2) the direct measurement of energetic costs of individual traits, and (3) quantification of the performance landscape. Obtaining data in these three areas is extremely difficult through the study of living organisms alone, and the use of robotic models can reveal unexpected effects. Controlling for all variables except for the length of a swimming flexible object reveals substantial non-linear effects that vary with stiffness. Quantification of the swimming performance surface reveals that there are two peaks with comparable efficiency, greatly complicating the inference of performance from morphology alone. Organisms and their ecological interactions are complex, and dissecting this complexity to understand the effects of individual traits is a grand challenge in ecology and evolutionary biology. Robotics has great promise as a “comparative method,” allowing better-controlled comparative studies to analyze the many interacting elements that make up complex behaviors, ecological interactions, and evolutionary histories.
Full-text available
Animal behaviour is remarkably sensitive to disruption by chemical pollution, with widespread implications for ecological and evolutionary processes in contaminated wildlife populations. However, conventional approaches applied to study the impacts of chemical pollutants on wildlife behaviour seldom address the complexity of natural environments in which contamination occurs. The aim of this review is to guide the rapidly developing field of behavioural ecotoxicology towards increased environmental realism, ecological complexity, and mechanistic understanding. We identify research areas in ecology that to date have been largely overlooked within behavioural ecotoxicology but which promise to yield valuable insights, including within-and among-individual variation, social networks and collective behaviour, and multi-stressor interactions. Further, we feature methodological and technological innovations that enable the collection of data on pollutant-induced behavioural changes at an unprecedented resolution and scale in the laboratory and the field. In an era of rapid environmental change, there is an urgent need to advance our understanding of the real-world impacts of chemical pollution on wildlife behaviour. This review therefore provides a roadmap of the major outstanding questions in behavioural ecotoxicology and highlights the need for increased cross-talk with other disciplines in order to find the answers.
Full-text available
Invasive species threaten biodiversity and ecosystem functioning. We develop an innovative experimental approach, integrating biologically inspired robotics, time-series analysis, and computer vision, to build a detailed profile of the effects of non-lethal stress on the ecology and evolution of mosquitofish (Gambusia holbrooki)—a global pest. We reveal that brief exposures to a robotic predator alter mosquitofish behavior, increasing fear and stress responses, and mitigate the impact of mosquitofish on native tadpoles (Litoria moorei) in a cause-and-effect fashion. Effects of predation risk from the robot carry over to routine activity and feeding rate of mosquitofish weeks after exposure, resulting in weight loss, variation in body shape, and reduction in the fertility of both sexes—impairing survival, reproduction, and ecological success. We capitalize on evolved responses of mosquitofish to reduce predation risk—neglected in biological control practices—and provide scientific foundations for widespread use of state-of-the-art robotics in ecology and evolution research.
Full-text available
Social learning is ubiquitous across the animal kingdom, where animals learn from group members about predators, foraging strategies, and so on. Despite its prevalence and adaptive benefits, our understanding of social learning is far from complete. Here, we study observational learning in zebrafish, a popular animal model in neuroscience. Toward fine control of experimental variables and high consistency across trials, we developed a novel robotics-based experimental test paradigm, in which a robotic replica demonstrated to live subjects the correct door to join a group of conspecifics. We performed two experimental conditions. In the individual training condition, subjects learned the correct door without the replica. In the social training condition, subjects observed the replica approaching both the incorrect door, to no effect, and the correct door, which would open after spending enough time close to it. During these observations, subjects could not actively follow the replica. Zebrafish increased their preference for the correct door over the course of 20 training sessions, but we failed to identify evidence of social learning, whereby we did not register significant differences in performance between the individual and social training conditions. These results suggest that zebrafish may not be able to learn a route by observation, although more research comparing robots to live demonstrators is needed to substantiate this claim.
Full-text available
Zebrafish (Danio rerio) constitutes a valuable experimental species for the study of the biological determinants of emotional responses, such as fear and anxiety. Fear-related test paradigms traditionally entail the interaction between focal subjects and live predators, which may show inconsistent behavior throughout the experiment. To address this technical challenge, robotic stimuli are now frequently integrated in behavioral studies, yielding repeatable, customizable, and controllable experimental conditions. While most of the research has focused on open-loop control where robotic stimuli are preprogrammed to execute a priori known actions, recent work has explored the possibility of two-way interactions between robotic stimuli and live subjects. Here, we demonstrate a “closed-loop control” system to investigate fear response of zebrafish in which the response of the robotic stimulus is determined in real-time through a finite-state Markov chain constructed from independent observations on the interactions between zebrafish and their predator. Specifically, we designed a 3D-printed robotic replica of the zebrafish allopatric predator red tiger Oscar fish (Astronotus ocellatus), instrumented to interact in real-time with live subjects. We investigated the role of closed-loop control in modulating fear response in zebrafish through the analysis of the focal fish ethogram and the information-theoretic quantification of the interaction between the subject and the replica. Our results indicate that closed-loop control elicits consistent fear response in zebrafish and that zebrafish quickly adjust their behavior to avoid the predator's attacks. The augmented degree of interactivity afforded by the Markov-chain-dependent actuation of the replica constitutes a fundamental advancement in the study of animal-robot interactions and offers a new means for the development of experimental paradigms to study fear.
Full-text available
Among territorial animals, several species are characterized by males showing the same initial behaviours towards both sexes, leading to significant chances of injuries against conspecifics. In this study, we investigated how visual stimuli exhibited by a female-mimicking robotic replica can be exploited by highly territorial Betta splendens males to discriminate males from females. In addition, we tested the effect of light stimuli, mimicking the colour pattern of a reproductive female, on the consistence of courtship displays in B. splendens males. The intensity of male behaviours used in both courtship and not-physical agonistic interactions (e.g. fin spreading and gill flaring) was not importantly modulated by different stimuli. Conversely, behavioural displays used specifically in male–female interactions significantly increased when the robotic replica colour pattern mimicked a reproductive female. Furthermore, male courtship behaviours obtained in response to the robotic replica exhibiting light stimuli were comparable with responses towards authentic conspecific females. Our biomimetic approach to establish animal–robot individual interaction can represent an advanced strategy for trait-based ecology investigation, a rapidly developing research field.
Full-text available
The effect of previous exposure to lateral sensory stimuli in shaping the response to subsequent symmetric stimuli represents an important overlooked issue in neuroethology, with special reference to arthropods. In this research, we investigated the hypothesis to ‘programme’ jumping escape direction as well as surveillance orientation in young and adult individuals of Locusta migratoria as an adaptive consequence of prior exposure to directional-biased predator approaches generated by a robotic leopard gecko representing Eublepharis macularius. The manipulation of the jumping escape direction was successfully achieved in young locusts, although young L. migratoria did not exhibit innately lateralized jumping escapes. Jumping escape direction was also successfully manipulated in adult locusts, which exhibited innate lateralized jumping escape at the individual level. The innate lateralization of each instar of L. migratoria in using a preferential eye during surveillance was not affected by prior lateralized exposure to the robotic gecko. Our results indicate a high plasticity of the escape motor outputs that are occurring almost in real time with the perceived stimuli, making them greatly adaptable and compliant to environmental changes in order to be effective and reliable. In addition, surveillance lateralization innately occurs at population level in each instar of L. migratoria. Therefore, its low forgeability by environmental factors would avoid disorganization at swarm level and improve swarm coordination during group tasks. These findings are consistent with the fact that, as in vertebrates, in insects the right hemisphere is specialized in controlling fear and escape functions.
Full-text available
Living organisms are far superior to state-of-the-art robots as they have evolved a wide number of capabilities that far encompass our most advanced technologies. The merging of biological and artificial world, both physically and cognitively, represents a new trend in robotics that provides promising prospects to revolutionize the paradigms of conventional bio-inspired design as well as biological research. In this review, a comprehensive definition of animal–robot interactive technologies is given. They can be at animal level, by augmenting physical or mental capabilities through an integrated technology, or at group level, in which real animals interact with robotic conspecifics. Furthermore, an overview of the current state of the art and the recent trends in this novel context is provided. Bio-hybrid organisms represent a promising research area allowing us to understand how a biological apparatus (e.g. muscular and/or neural) works, thanks to the interaction with the integrated technologies. Furthermore, by using artificial agents, it is possible to shed light on social behaviours characterizing mixed societies. The robots can be used to manipulate groups of living organisms to understand self-organization and the evolution of cooperative behaviour and communication.
Full-text available
In the field of animal behavior, effective methods to apprehend causal relationships that underlie the interactions between animals are in dire need. How to identify a leader in a group of social animals or quantify the mutual response of predator and prey are exemplary questions that would benefit from an improved understanding of causality. Information theory offers a potent framework to objectively infer cause-and-effect relationships from raw experimental data, in the form of behavioral observations or individual trajectory tracks. In this targeted review, we summarize recent advances in the application of the information-theoretic concept of transfer entropy to animal interactions. First, we offer an introduction to the theory of transfer entropy, keeping a balance between fundamentals and practical implementation. Then, we focus on animal-robot experiments as a means for the validation of the use of transfer entropy to measure causal relationships. We explore a test battery of robotics-based protocols designed for studying zebrafish social behavior and fear response. Grounded in experimental evidence, we demonstrate the potential of transfer entropy to assist in the detection and quantification of causal relationships in animal interactions. The proposed robotics-based platforms offer versatile, controllable, and customizable stimuli to generate a priori known cause-and-effect relationships, which would not be feasible with live stimuli. We conclude the paper with an outlook on possible applications of transfer entropy to study group behavior and clarify the determinants of leadership in social animals.
The possibility of regulating the behavior of live animals using biologically-inspired robots has attracted the interest of biologists and engineers for over 25 years. From early work on insects to recent endeavors on mammals, we have witnessed fascinating applications that have pushed forward our understanding of animal behavior along new directions. Despite significant progress, most of the research has focused on open-loop control systems, in which robots execute predetermined actions, independent of the animal behavior. In this article, we integrate mathematical modeling of social behavior toward the design of realistic feedback laws for robots to interact with a live animal. In particular, we leverage recent advancements in data-driven modeling of zebrafish behavior. Ultimately, we establish a novel robotic platform that allows real-time actuation of a biologically-inspired three-dimensionally printed zebrafish replica to implement model-based control of animal behavior. We demonstrate our approach through a series of experiments, designed to elucidate the appraisal of the replica by live subjects with respect to conspecifics and to quantify the biological value of closed-loop control.
Adjusting behaviour can be crucial to prey surviving a predator encounter. How any one individual modifies their behaviour in response to predation risk might be affected by their previous experience with danger and their own vulnerability. Using western mosquitofish, we examined how boldness in different contexts was affected by an individual's recent experience with predation risk. Individuals were repeatedly chased by a largemouth bass model and encountered alarm cue to mimic conditions of high risk (cues twice on 2 days), low risk (cues twice on 1 day), or no risk (water only). We then measured boldness and avoidance behaviour under three different contexts: in a novel tank, with a shoal of conspecifics, and with alarm cues and a model predator. We found that how recent experiences influenced boldness in a novel tank depended on body size. Smaller fish from the no and low risk treatments were more likely to emerge from a shelter into a novel environment than larger individuals. When individuals had recently experienced high levels of risk however, this pattern was reversed. We also found that individuals who had experienced any recent risk (low and high)were more likely to leave the safety of a shoal and approach a novel object compared to individuals who had not experienced any recent danger. Avoidance behaviour across the three assays was not affected by recent experiences but was affected by body size to varying degrees. For example, larger fish were more likely to stay in the plants, away from the cues of predation compared to smaller fish. Overall, our results suggest that how recent experiences with risk influence subsequent behaviour can depend on a variety of interacting factors including the intensity of recent experiences, the particular behaviour examined, and an individual's body size.
Swimming in schools affords several advantages for fish, including enhanced ability to escape from predators, searching for food, and finding correct migratory routes. However, the role of hydrodynamics in coordinated swimming is still not fully understood due to a lack of data-driven approaches to disentangle causes from effects. In an effort to elucidate the mechanisms underlying fish schooling, we propose an empirical study that integrates information theory and experimental biology. We studied the interactions between an actively pitching airfoil and a fish swimming in a flow. The pitching frequency of the airfoil was varied randomly over time, eliciting an information-rich interaction between the airfoil and the fish. Within an information-theoretic framework, we examined the information content of fish tail beating and information transfer from the airfoil to the fish. The proposed framework may help improve our understanding of the role of hydrodynamics in fish swimming, thereby supporting hypothesis-driven studies on the hydrodynamic advantages of fish schooling.