Research Article

International Journal of Advanced

Robotic Systems
May-June 2017: 1–12
Design of a modular robotic system that ª The Author(s) 2017
DOI: 10.1177/1729881417706628
mimics small fish locomotion and body journals.sagepub.com/home/arx

movements for ethological studies

Frank Bonnet1, Leo Cazenille2, Axel Séguret2, Alexey Gribovskiy1,

Bertrand Collignon2, José Halloy2 and Francesco Mondada1

Abstract
Robotic animals are nowadays developed for various types of research, such as bioinspired robotics, biomimetics, and
animal behavioral studies. The design of these robots poses great challenges as they often have to achieve very high-level
performances in terms of locomotion, size, and visual aspect. We developed a robotic system for direct underwater
interactions with small fish species. This robotic platform is composed of two subsystems: a miniature wheeled mobile
robot that can achieve complex locomotion patterns and a robotic fish lure that is able to beat its soft caudal peduncle to
generate fish-like body movements. The two subsystems are coupled with magnets that allow the robotic lure to reach
very high speeds and accelerations, thanks to the mobile robot. We used zebrafish (Danio rerio) to model small fish
locomotion patterns and construct a controller for the motion of our robotic system. We have demonstrated that the
designed system is able to achieve the same types of motion patterns as the zebrafish while mimicking the body move-
ments of the fish. These results define new standards for robotic fish lures and bring to the field of fish–robot interaction a
new tool for ethological studies.

Keywords
Biomimetic robots, animal–robot interaction, underwater robotics, soft robotics, zebrafish

Date received: 24 April 2016; accepted: 3 August 2016

Topic: Special Issue - Soft Robotics Interacting with the Real World
Topic Editor: Matteo Cianchetti

Introduction be found. For instance, in earlier studies,6–8 zebrafish

(Danio rerio) response to a robotic fish was observed. The
It is common nowadays to use robotic devices to study the
robotic fish, which had the same aspect ratio as the zebra-
behavior of animals, both in their natural environments1
fish, was attached to a moving device on top of a tank and
and in laboratory research. For instance, robotic platforms its speed could be varied, as well as the tail beating and its
were socially integrated into a society of cockroaches,
where they emitted pheromones and reproduced the
motion behavior of cockroaches2; they were also intro- 1
Robotic Systems Laboratory, School of Engineering, Ecole Polytechnique
duced into a group of chicks using a filial imprinting Fédérale de Lausanne, Lausanne, Switzerland
mechanism3,4; or designed to mimic the honeybee dance 2
Paris Interdisciplinary Energy Research Institute, University Paris Diderot
communication system by generating wing oscillations VII, Paris, France
and heat.5
Corresponding author:
The collective behavior of the fish, in particular, raised Frank Bonnet, Ecole Polytechnique Federale de Lausanne, EPFL STI IMT
the interest of scientists, and several examples of automated LSRO, ME B3 30 (Batiment ME), Lausanne 1020, Switzerland.
lures designed to interact with fish underwater can already Email: [email protected]

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(http://www.creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
2 International Journal of Advanced Robotic Systems

coloration. A study by Phamduy et al.9 measured the pre- guppies, golden shiners, or sticklebacks. A controller based
ference of fertile female bluefin killifish (Lucania goodei) on data extracted from experimental zebrafish trajectories
for a robotic replica whose aspect ratio, body size, motion is implemented on the robotic device to reproduce their
pattern, and color were inspired by an adult male killifish locomotion patterns. Finally, we show the matching of the
using a robotic platform specifically designed to simulate locomotion pattern of our robot with that of the zebrafish
the typical courtship behavior observed in male killifish. In using the information retrieved by a vision-based tracking
the studies by Marras and colleagues,10–14 a passive lure system to qualify our implementation.
attached to a support was moved using a mobile robot The capabilities of this new robotic platform will help
below the aquarium and controlled using a tracking soft- scientists to better understand the type of interactions
ware. In a study by Worm et al.,15 a mobile robot was also between small species of fish that lead to the formation
used to move a robotic lure, which emitted electrical play- of shoaling and to collective choices.
back signals to attract the weakly electric fish (Mormyrus
rume). While these different studies demonstrated the
potential to develop artificial devices that are able to inter- Hardware design
act with fish, there is no solution involving an active lure The robotic platform presented in this research is called
with size close to the size of the fish that can reproduce the Fish-Control-Actuator-Sensor-Unit or Fish-CASU. The
pattern motion of the fish and autonomously move in an Fish-CASU is composed of two modules: a wheeled
aquarium to integrate a shoal of fish. mobile robot, the FishBot, that is able to achieve high
With regard to the bioinspired research applied on fish- speeds and accelerations while performing local obstacle
like robots, there is a trend to apply new types of actuators avoidance using infrared (IR) proximity sensors and an
for the development of small-scale biomimetic fish actuated bio-inspired fish lure, the RiBot, that is equipped
robots.16–20 Even if this allows high levels of performance with a beating caudal peduncle to reproduce fish body
to be achieved in terms of size, tail beating frequencies, or movements. The two modules are coupled with magnets
energy consumption, these actuators have several draw- that allow the RiBot to follow a shoal of fish swimming at
backs, such as temperature sensitivity for shape-memory high speed while moving its body similarly to the fish.
alloys and specific liquid environment requirements for The hardware design of the two modules will be described
electroactive polymers which can pose challenges to the separately in the next subsections, followed by the experi-
design of an autonomous and miniature robotic fish. More- mental setup used to perform the biological experiments
over, it is sometimes difficult to embed the electronics into on zebrafish.
the devices due to the size of the components required to
drive the different actuators; so either the fish need to be
rather big or the electronic need to be partially external FishBot, the high-speed miniature mobile robot
which reduce the autonomy of the system. The design of the FishBot was driven by the following
Regarding the embodiment of fish locomotion on a requirements: the robot should achieve linear speed and
robotic device to reproduce fish behavior, few models of acceleration similar to the zebrafish, and it should allow
fish locomotion can be found in the literature. For instance, multi-robot, long duration experiments. Thus, its width and
in a study by Mwaffo et al.,21 fish locomotion is modeled length should not exceed that of zebrafish to allow the
using a jump persistent turning walker model motivated by shoaling of Fish-CASUs, and it should be continuously
the sudden and drastic changes in zebrafish locomotion in powered. Some prototyped versions of the FishBot were
the form of large deviations in turn rate. Zienkiewicz already introduced in literature,23,24 but we still have
et al.22 used a stochastic model to reproduce zebrafish loco- improved the maneuverability of the device to fit the spec-
motion in a confined environment. While these models ifications to conduct this study.
accurately reproduce the motion of fish, the translation of A differential drive configuration with two indepen-
their mathematical expression in concrete commands for a dently motorized wheels was adopted for the locomotion.
robot was not validated on a real system. It is enough to mimic the zebrafish locomotion in 2-D as
In this article, we present the combination of two robotic they usually only move with a tangential speed and no
modules: a highly maneuverable wheeled robot, prototyped lateral speed. In terms of rotational speed, due to the differ-
versions of which were presented in the studies by Bonnet ential drive configuration and the small distance between
and colleagues,23,24 and an actuated robotic fish lure the two wheels, the FishBot can also have very high capa-
designed using a biomimetic approach. It allows robots to cities which are enough to mimick the rotational motions of
navigate inside species of small fish shoals at very high zebrafish. Two DC motors (Maxon Motor AG, Sachseln,
speeds and with multiple robotic agents while reproducing Switzerland) are used to drive the two robot wheels through
specific fish body movements which has not been achieved a worm gear transmission of 16:1 reduction ratio.
in related studies yet. In this study, we use the zebrafish as a The current version of the FishBot is presented in
model animal, but this system could be also used to study Figure 1. The length of the mobile robot is 55 mm, the
the behavior of other types of small freshwater fish, such as width 22 mm, and the height 53 mm. The mass of the
Bonnet et al. 3

motors at voltages higher than the manufacturer-specified

nominal voltage, thus providing a higher torque for a short
period of time.

RiBot, the active soft robotic lure

Based on related studies on zebrafish–robot interactions,6–8
the lure acceptance among the zebrafish shoal can be
increased when the lure beats its tail. It is also known that
zebrafish can bend their caudal peduncle up to very high
angle (up to p rad) with high angular velocities (up to
4.2 rad s1).21 Therefore, we designed an actuated lure
able to emit such tail beating stimulus. While the beating
Figure 1. Current version of the FishBot used for mimicking fish of the tail could have been generated passively using the
motion patterns. (a) Magnets to couple the FishBot with the lure water flow generated by the motion of the lure, we devel-
module; (b) electric brushes to retrieve the power from the oped an actuator inside the lure to actively control the tail
positive conductive plate; (c) microcontroller dsPIC33f128; (d) beating. This actuator placed inside the lure allows us to
supercapacitors that store power if the contact with the plates is determine the frequency and amplitude of the tail beating
lost; (e) Bluetooth antenna; (f) Maxon DC motor; (g) infrared
independently from the motion of the lure. This device has
proximity sensors; (h) wheel; (i) electric brushes to retrieve the
power from the ground connected conductive plate. to be wireless and waterproof, have a low power con-
sumption, and should fit inside a small lure since zebrafish
FishBot is 80 g. The power supply is done through electric rarely exceeds 45 mm in length.
cables (brushes) that slip against two conductive plates The design of the RiBot involved the use of a printed
situated under the aquarium floor (positive voltage) and circuit board (PCB) on which all the electronic components
on the support on which the robot moves (ground; Figure 4). are soldered. As PCB also has very interesting mechanical
Supercapacitors are used in order to store energy in case the properties, the PCB also forms the skeleton of the device,
brushes are not in contact with the conductive plates for a combining in this way mechanical and electronic structures
couple of seconds to allow the continuous powering of the to reduce the size and weight of the system. The RiBot is
robot. Six IR proximity sensors are installed on the front equipped with a rechargeable lithium polymer (LiPo) bat-
and back of the robot in order to avoid obstacles, other tery of 40 mAh capacity with a size of 12  16  5 (mm).
FishBots, or the border of the arena. A Bluetooth antenna An IR sensor is placed on the back of the RiBot, and an IR
is used for wireless communication to control the robot. signal can be sent from any direction to control the device
An ASEBA virtual machine runs onboard each mobile underwater. For the actuation of the caudal peduncle, we
robot. ASEBA25 is an event-based architecture for the real- chose the microstep gear motor MF03G (Seiko Precision
time distributed control of robots. The ASEBA communi- Inc., Matsudo, Japan). This stepper motor has enough tor-
cation protocol allows multiple robots to connect on the que to drive the caudal peduncle underwater.
same network. This allows them to share global events that Figure 2 shows the manufacturing steps of the RiBot.
can be sent via a high-level control application or by the The motor with the caudal peduncle docked on its axis is
robots themselves. ASEBA, integrated with D-Bus, allows fixed inside a 3-D part, called the ring as it has an elliptical
access to each robot using high-level programming lan- external shape, that is glued onto the PCB. A LiPo battery
guages (such as Cþþ, Python, or ROS) via a software hub of 40 mAh capacity is soldered onto the eye of the RiBot,
called Medulla.26 which will then be used to recharge the device. In order to
In order to increase the capabilities of the robot to isolate the actuator and the caudal peduncle from water and
mimick the locomotion of zebrafish in terms of speed and to create a soft skin that can mimic the tail of the zebrafish,
acceleration, a cascade controller (Figure 8) is implemen- an undercut of the tail was made in 3-D printing and was
ted for the low-level control of each motor. Each motor dipped into liquid latex. The skin created was unmolded
torque is controlled by a proportional–integral (PI) control- and attached on the ring using silicone. The caudal fin is
ler running at 1 KHz, which is itself controlled in speed by a also made of latex and is prepared separately from the tail
proportional–integral–derivative controller or in position using another mold, and it is glued onto the tail using latex.
by a proportional–derivative controller using the motor Polyurethane is used to isolate the electronics from the
encoders. An infinite impulse response (IIR) filter with the water. The PCB is placed inside a mold with the desired
same time constant as that of the motor is run on the micro- undercut of the RiBot, and the eyes are used as a reference
controller. By precisely controlling the current inside each inside the mold. Liquid polyurethane is injected inside the
motor, based on the estimated power dissipation from the mold from the tail. The polyurethane coats the entire PCB
IIR filter, we are able to ensure that we never overheat up to the ring and hardens inside the mold. For the shape
the motor. Such control architecture enables us to use the design of the molds used to manufacture the latex tail and
4 International Journal of Advanced Robotic Systems

Figure 2. Process of the RiBot manufacturing. (a) The stepper

motor and the eyes are soldered onto the PCB. (b) The battery is
placed and connected to the eyes. The caudal peduncle is fixed to
the stepper motor axis and a ring is placed on the stepper motor.
(c) A latex socket is placed and glued using silicone onto the ring
to isolate the back of the fish from water and reproduce the
caudal fin skin. (d) The RiBot is put inside a mold and polyurethane
is injected into it. The polyurethane covers all the rest of the body Figure 3. The RiBot compared with one of our zebrafish. This
so that the device is totally waterproof. PCB: printed circuit figure shows the 1.8 ratio between the lure and a real zebrafish.
board. The RiBot is mounted on a base composed of a tiny carbon stick
and an iron plate. Magnets are attached to the iron plate to
the polyurethane body, we used surfaces extracted from a magnetically couple the RiBot with the FishBot. Here, we used
3-D scan of a zebrafish. The surface of the skin was zebrafish as a model animal, but the RiBot could also be adapted
extracted and increased by the proper ratio to match the to study the behavior of other small freshwater fish, thanks to its
small size.
size of the PCB. After this process, the RiBot is totally
isolated from water.
A comparison of a prototype version of the RiBot with
one of our zebrafish is shown in Figure 3. The RiBot has a
mass of 9.0 g and sinks beneath the water surface due to its
density. It was not possible with the selected components to
obtain a robot with the same scale as the zebrafish, thus the
current design has a 1.8 ratio compared to the zebrafish. In
terms of autonomy, we performed a 24-minute test with the
caudal peduncle of the RiBot moving continuously using
the embedded LiPo battery; however, upon reducing the
use of the actuator, the RiBot remained turned on for more
than 1 h underwater. This allowed us to perform hour-long
experiments with the Fish-CASU. Both polyurethane and
latex surfaces can be painted in order to mimic the fish Figure 4. Experimental setup used during the experiments. (a)
Camera used to track the lures and the zebrafish; (b) infrared
colors, as it was done in the study by Abaid et al.6
emitter; (c) Raspberry Pi; (d) RiBot inside the aquarium linked to a
FishBot through magnetic coupling; (e) living zebrafish; (f) aqua-
rium of 1000  1000  250 (mm); (g) water layer of 60 mm depth;
The experimental setup (h) conductive plates to power the mobile robot; (i) FishBot
The experimental setup used for the experiments consists moving under the aquarium; (j) the computer that processes the
of an aquarium of 1000  1000  200 (mm) covered on camera frames and remotely controls the robots via Bluetooth
and infrared RC5 protocol.
the inside with white Teflon sheets (Figure 4). These
sheets are installed in order to avoid reflections on the
glass and to provide a smooth surface for the motion of AG, Germany) with a maximum resolution of 2048 
the lure module inside the aquarium. The structure is sup- 2048 (pixels) and equipped with low distortion lenses
ported by a rigid structural aluminum framing system, CF12.5HA-1 (Fujinon, Tokyo, Japan) grabs frames that
which is exposed to diffused light in order to reproduce are then processed for tracking purposes. An IR light-
daylight. The tank is filled with water up to a level of 60 emitting diode is placed on top of the setup to control the
mm. The FishBots are powered using two conductive RiBots. It is connected to the output of a Raspberry PI that
plates, one glued onto the bottom of the aquarium and one is interfaced with the main desktop computer. The RC5 IR
onto the support on which the robot moves. An overhead signals are generated using the Linux IR Remote Control
acA2040-25gm monochrome GigE CCD camera (Basler (LIRC) library on the Raspberry PI.
Bonnet et al. 5

Table 1. Comparison of the size and linear speed of the fish robots presented in other similar studies.

System Length (mm) Height (mm) Width (mm) Linear speed (mm s1)
Abaid, 2012 150 48 26 Not moving
Aureli, 2012 90 45 35 8–12
Butail, 2014 117 48 26 40
RiBot 75 17 10 300
Zebrafish 45 10 5 450

Comparison of the Fish-CASU with related work on other robotic fish designs while respecting more the
fish–robot interaction shape ratio of a zebrafish. Finally, in terms of linear
speed, in a study by Aureli et al.,7 the lure is moving
The designed system described in this section brings sev- autonomously underwater and thus its linear speed is
eral new assets compared to existing published solutions: quite small. In a study by Butail et al., 8 the lure is
the multi-robot long duration experiments and the small attached to a robotic arm that allows it to move with
size of the robotic lure that is in direct contact with the fish. speeds up to 40 mm s1. Thanks to the coupling with the
FishBot, the RiBot is able to move with much higher
Multi-robot experiments. In order to reproduce the motion of speeds than the other existing solutions.
a fish shoal using robotic devices, the robots have to move
very close to each other, thus should have very small size.
With the proposed solution, we can run multi-robot experi-
Control and tracking software
ments with a large number of robots moving in every pos-
sible directions that cannot be achieved by the methods In order to retrieve the position of the agents (fish and Fish-
presented in the studies by Faria et al.11 and Phamduy CASUs) during an experiment and to control the Fish-
et al.9 However, these types of experiments are also possi- CASU, we have implemented a vision-based tracking in
ble with the approach of the studies by Landgraf and col- our control software that processes the frames grabbed by
leagues12–14 that also use a differential drive mobile robot the upper camera of the experimental setup. A background
to move a dummy fish underwater through magnetic cou- subtraction is performed on each frame to extract the mov-
pling. However, comparing the size of the mobile robots, ing objects. Then we detect the head of each agents using a
Landgraf et al.12,13 used a mobile robot of 70 mm of length corner detection methods that is described in a study by Shi
and 70 mm of width, while Swain et al.14 used the MiaBot27 and Tomasi.28 The poses of the RiBots are used to estimate
that has a length and a width of 75 mm, thus the distance the poses of the FishBots for their closed-loop control. The
between two dummy fish will be of 70 mm in minimum beating of the RiBot caudal peduncle is also controlled by
for the first case and 75 mm for the second case. With the this application.
FishBot, thanks to it width of 22 mm, it is possible to have In terms of global architecture (Figure 5), the control of
two lures moving at a distance of 22 mm, thus much closer the FishBots motion is done through ASEBA events that
to natural inter-fish distances. This is one of the main are sent from the main application and that contain the
advantages of the FishBot compared to other robots. parameters for the locomotion. The behaviors that will be
Moreover, the robots are also equipped with IR proximity described in the next section are implemented onboard
sensors that can be used to avoid any collisions between each FishBot. Thanks to the event-based protocol, the
the FishBots. FishBots are able to emit events in case of obstacle pres-
ence or powering issues. The detection of obstacles is
Long duration experiments. Another advantage is the contin- done using the IR proximity sensors: when an IR signal
uous powering of the system. Indeed, the robotic devices received by the sensors is higher than a certain threshold,
used in the studies by Landgraf and colleagues12–14 are the FishBot avoid the obstacle by turning with a certain
powered using onboard batteries. With our proposed amount of time in the opposite direction of the received
design, the powering of the mobile robot is continuous, signal. It allows the FishBot to avoid the collisions with
which offer experiments that can last up to several days the walls as well as other FishBots.
if using a passive lure such as the lures proposed in the A Raspberry PI, on which LIRC library is run, is con-
studies by Landgraf and colleagues12–14 or RiBot without nected to the same network as the main computer, and
the continuous actuation of the tail. RC5 signals are generated on an output pin connected to
an IR emitter in order to control the RiBots. The IR signal
Small-size actuated lure. Table 1 shows a comparison of the is broadcasted on the whole aquarium and received by all
RiBot with other actuated robotic lures designed in other the RiBots.
studies on fish–robot interaction. In terms of size, we The tracking and control application was also used to
have smaller length, width, and height compared with analyze the locomotion of the agents. Using the positions of
6 International Journal of Advanced Robotic Systems

Figure 6. (a) The trajectory of a zebrafish swimming in the

experimental tank for 1 h. The color of the trajectory indicates
the linear speed of the fish at a given position (Dt ¼ 1 s). The fish is
mainly observed along the wall of the tank and decreases its speed
in the corners of the tank, while higher speed values are measured
in the center of the aquarium. (b) Zone delimitation in a tank of L
(100 cm) and W (100 cm) for the three different types of fish
Figure 5. Architecture of the Fish-CASU vision-based closed- behavior extraction: corners (green) that are delimited by the
loop control. The high-level control-software tracks each RiBot right angle walls of the tank and virtual lines at a distance d (15 cm)
and fish using frames grabbed by a camera. The same software is from the walls, edges (blue) delimited by the setup wall and the
used to control the FishBots motion as well as the RiBots body same virtual line, and the center (red). The motion patterns of the
movements. Each RiBot and each FishBot can be controlled fish were classified depending on the presence zone.
individually. For the FishBots, the desired speed of each wheel is
sent through a serial connection using Medulla integrated with D-
Bus on the ASEBA network, where all the FishBots are connected pH level was maintained at 7.5, and nitrites (NO2) were
and can receive or emit events. For the RiBots, the communica- below 0.3 mg l1.
tion is only one way, where the IR RC5 signal is broadcasted to
control stepper motors that drive the caudal peduncles.
Zebrafish locomotion analysis
Zebrafish were individually transferred from their housing
the agents moving inside of the aquarium, retrieved by the
aquarium to the experimental aquarium to extract the char-
tracking, we could compute the speeds knowing that the
acteristics of their trajectories over time. We recorded, at
tracking was done at 15 frames per second. We computed
15 frames per second, the position of each fish swimming
the linear and angular speed using the position difference
alone in the tank for 1 h. Thanks to this tracking, we
between three frames instead of two frames in order to
retrieved the trajectory of each fish and computed their
average the noise given by the tracking.
speed and acceleration during the entire experiment. An
example of a trajectory is given in Figure 6(a) and high-
Locomotion pattern embodiment lights that the fish were mainly swimming along the walls
of the tank, as also shown by Zienkiewicz et al.22 We
Ethics statement could also identify three speed patterns of the fish accord-
The experiments performed in this study were conducted ing to their location in the experimental tank. Indeed, their
under the authorization No 2778 delivered by the Depart- linear speed decreases in the corners of the tank when they
ment of Consumer and Veterinary of the Canton de Vaud change their direction, while the zebrafish swim at a
(Switzerland) and the Buffon Ethical Committee (regis- higher speed along the edges of the tank. Finally, we
tered to the French National Ethical Committee for Animal measured the highest linear speed values in the middle
Experiments #40) after submission to the French and Swiss of the aquarium. Thus, we divided the experimental tank
state ethical board for animal experiments. into different zones: corners, edges, and the center (Figure
6(b)). The distance d delimiting the edges and the corners
was estimated at 15 cm.
Animals and housing As our goal is to develop a modular robotic system that
To characterize the locomotion pattern of zebrafish, we mimics fish locomotion, we also performed a more detailed
observed 10 adult wild-type zebrafish (Danio rerio) AB analysis of the characteristic motion pattern of the fish. The
strain acquired from Institut Curie (Paris, France). The speed of the fish can be decomposed into successive cycles,
fish were 18 months old at the time of the experiments. starting with a sharp acceleration followed by a slow decel-
We kept the fish under laboratory conditions, 27  C, eration until the next tail beat (Figure 7(a) and– (b), left).
500 S salinity, with a 10:14 day–night cycle. The fish With regard to the direction of the fish, the changes in
were reared in 55 l tanks and fed two times a day (Special orientation, identified by a high angular speed, are mainly
Diets Services SDS-400 Scientific Fish Food). The water detected at the beginning of the cycle (Figure 7(c), left).
Bonnet et al. 7

Figure 7. Left: Example of the characteristic swimming behavior of a zebrafish magnified for 3 s. Each colored segment represents the
linear speed (a), acceleration (b), and angular speed (c) of the fish measured at a given position (Dt ¼ 1/15 s). The linear speed of the fish
can be decomposed into cycles that start by a sharp acceleration (hot colors) followed by a longer deceleration (cold colors). The
angular speed highlights that changes of orientation occur mainly at the start of the cycles. Right: Finite-state machine implemented on
the control layer of FishBots. A motion cycle is started when an event is received from the high-level control application. The
parameters contained on the event are used in the different steps: Dy is the angle difference between the current orientation of the
RiBot and the orientation needed to reach the next target, Pt is the target position to accelerate, and Vt is the target linear forward
speed. If an obstacle or another robot is detected during a cycle, a simple obstacle avoidance behavior is implemented to avoid it until a
new event is received.

Thus, the zebrafish usually move by following a sequence

of three steps:

 Step 1: Orientation. Strong caudal peduncle bending

to reorient and start the propulsion toward the next
goal.
 Step 2: Acceleration. High linear acceleration fol-
lowing the given caudal peduncle beating to
reorient.
 Step 3: Relaxation. Tail beating stops and the fish
starts sliding into water with its linear speed
decreasing. Figure 8. Control architecture of the FishBot motors. A PI
current controller is running at 1 KHz. The motor encoders
retrieve the position of the motor that is used in a PID speed
control loop and PD position control loop at 200 Hz. The user
Implementation on the Fish-CASU can choose between position or speed control of the motor.
We implemented a finite-state machine into the robot con-
troller in order to execute the same locomotion sequence as move forward while the other wheel remained stationary.
the zebrafish (Figure 7, right). An event containing the ID This generates a rotation of the robot with a small linear
of the FishBot is emitted with a parameterized frequency speed, as zebrafish almost never have a null linear speed
from the control application and sent to the ASEBA net- even while turning.
work on which all the FishBots are connected. When an For step 2, the acceleration, as we have no direct control
event is received by a FishBot, it starts executing the finite- over the acceleration (Figure 8), we measured the accelera-
state machine with the three locomotion steps described: tion obtained using the position control in cascade with the
orientation, acceleration, and relaxation. In case of an speed and current control. By using the cascade controller,
obstacle detected by the IR proximity sensors, the execu- the motors are protected against overheating or too high
tion of the locomotion is stopped and the robot starts avoid- speeds, and we can adjust the position controller gains to
ing the obstacle, and then return to the relaxation state obtain very high acceleration. The measures are shown in
where it waits for a new event. Figure 9 (left); we can see that the increase in acceleration
For step 1, the orientation, we used the wheels position is linear with the target position Pt ordered up to 6 cm,
control to reorient the robot toward the target with an orien- where the maximum FishBot acceleration of 1.3 m s2 of
tation difference of Dy. We used one wheel of the robot to average was achieved. After that, the motors entered in
8 International Journal of Advanced Robotic Systems

Figure 9. Left: Fish-CASU acceleration measured as a function of the position to reach using a PD position control. Right: Fish-CASU
speed measured as a function of the speed to reach.

Table 2. The parameters of the locomotion of the Fish-CASU for We recorded, at 15 frames per second, the position of the
the three zones of the experimental setup.a Fish-CASU moving alone in the tank for 1 h using the same
Zone F (Hz) Pt (Step 2) (cm) Vt (Step 3) (cm s1)
experimental setup and tracking software used to analyze
zebrafish movements. The Fish-CASU was programmed to
Corners 5 2 7 follow a trajectory similar to the zebrafish by following the
Edges 8 2.7 12 walls and, from time to time, moving rapidly in the center
Center 7 4.2 17
of the tank.
F: the frequency at which the ASEBA events are sent to start a motion
cycle; Pt: the target position to mimic the acceleration phase; Vt: the
starting linear speed for the relaxation phase.
Results and discussion
Figure 9 shows the tracking result of the FishBot for
saturation due to the limit of the system. The linear curve of
different speeds and accelerations. In respect to the existing
this figure was used to match the acceleration of the Fish-
literature, the FishBot has very high capacities in terms
Bot with the acceleration of the zebrafish. We have also
of acceleration (up to 1.3 m s2), linear speed (up to
characterized the limits of the system in terms of speed
30 cm s1), and rotational speed (more than 18 rad s1).
(Figure 9, right). We can observe that for speeds of up to
We also tracked the tip of the caudal fin of the RiBot in
20 cm s1, the response is relatively smooth and stable. For
order to measure the tail beating frequencies and ampli-
higher speeds, the robot takes more time to reach the
tudes (Figure 10). The RiBot has limited amplitude and
desired command, and over 30 cm s1, the limit of the
frequency capabilities that restrict it from reproducing
system is achieved. Finally, in terms of angular speed, the
some of the extreme body movements of zebrafish, which
FishBot alone is able to achieve angular speed of up to 30
can bend their caudal peduncle more than p=2 rad (90 )
rad s1. However, the magnetic coupling between the
with an angular speed over 4 rad s1; with the maximal
RiBot and the FishBot was lost over speeds of 18 rad s1.
beating amplitude of the RiBot is 0.52 rad (30 ), and the
Finally, for step 3, the relaxation, we used the speed
maximal angular speed is 3.6 rad s1 but for very low
controller to generate a linear forward speed Vt of the robot,
amplitudes. However, the RiBot is able to reproduce the
which decreases over time at a rate of 1 cm every 100 ms
average turn rate of the zebrafish (approximately 2.8
(10 cm s1), thereby reproducing the deceleration of the
rad s121) for an amplitude of 0.32 rad (18.35 ), as it can
zebrafish.
be seen in Figure 10.
The distribution of the linear speed measured in the
three zones of the tank (corners, edges, center) is shown
Experiments with the Fish-CASU in Figure 11 for the Fish-CASU and the zebrafish. As sug-
Regarding the matching of locomotion, we tuned the para- gested by preliminary observations, the fish swim with a
meters of our fish locomotion behavior implemented on the higher speed near the edges (0.054 + 0.032 m s1) than in
FishBot to match the data extracted from the zebrafish the corners (0.066 + 0.034 m s1) of the tank and even
locomotion patterns. The final parameters of the Fish- faster in the center of the aquarium (0.073 + 0.031 m s1).
CASU are shown in Table 2. These parameters were tuned By adjusting the parameters of the controller, we were able
in order to match the average linear speeds and, seconda- to reproduce similar distributions between the Fish-CASU
rily, the locomotion sequence curve in terms of speed and and the zebrafish. However, the fitting quality of the speed
acceleration (Figures 11–13). distribution is lower in the case of the Fish-CASU for
Bonnet et al. 9

Figure 10. Left: Top view of the RiBot beating its caudal peduncle. Right: Results of the tracking of the caudal fin for different beating
amplitudes and frequencies.

Figure 11. Speed density for the motion of an agent, Fish-CASU and zebrafish over the span of an hour in the three different zones of
the aquarium: corners, edges, and center.

Figure 12. The typical speed and acceleration of an agent, Fish-CASU and zebrafish, movements in a corner of the aquarium recorded
over 6 s.

motions along the edges and in the center, which is not the high average speeds, the Fish-CASU is less regular than the
case for the zebrafish. This can be explained by the fact that zebrafish due to the friction of the lure module, the slipping
for high accelerations that are applied in order to achieve of the FishBot wheels, and the inertia of the system.
10 International Journal of Advanced Robotic Systems

Figure 13. The average agents movement sequence for movement along the edges of the experimental tank. Linear speed (a and c) and
acceleration (b and d) of the Fish-CASU and the zebrafish during the speed cycle. The cycle begins with a short acceleration (2/15 s) until
the agent reaches its maximum speed, then it decelerates by gliding in the water until the next tail beat.

In addition, with the developed controller, the Fish-CASU which is composed of two units: a FishBot, a fast-
was able to reproduce the sinusoidal shape of the speed and moving wheeled robot with a small width and a continu-
acceleration of the fish (Figure 12). This is made possible by ous powering that is able to move a lure inside an
fitting the movement patterns of the fish with the movement aquarium, and a RiBot, a robotic lure that is only 1.8 times
patterns of the robots. By identifying successive local minima bigger than a zebrafish (Danio rerio), a model animal for
in the linear speed of the fish, we computed the average speed behavior studies. This robotic platform is modular in a
cycle of the fish as well as the corresponding acceleration. As way that the FishBot can be coupled with different types
previously shown, the sequence begins with a short accelera- of lure designed to interact with different small species of
tion during 2/15 s, and once the fish has reached its maximum fish. In parallel, the RiBot is a lure able to emit tail beating
speed, it glides and slows down to return to its initial speed stimuli, thanks to its actuated tail.
(Figure 13(c) and (d)). The Fish-CASU can mimic relatively The Fish-CASU can achieve very high acceleration and
well the motion pattern of the zebrafish. The mean speed linear speed, similar to the zebrafish. The local obstacle
obtained, as already shown in Figure 11, is similar. The initial avoidance behavior and the continuous powering of the
speed is smaller in the case of the Fish-CASU than for the fish; device allowed us to run very long duration experiments
this is explained by the friction of the lure module on the without any human intervention during the experiment.
aquarium floor. The standard deviations are relatively similar, While the tail beating of the RiBot does not recreate the
which shows that the robotic platform can reproduce well the exact same hydrodynamic patterns that fish generate, it can
typical locomotion patterns of a zebrafish. reproduce the average tail beating range of the zebrafish
that has been shown to be an attractive stimulus.6–8
We have studied the locomotion behavior of zebrafish in
Conclusion a rectangular tank in order to establish a controller for the
In this article, we presented a modular robotic fish locomotion of the Fish-CASUs. First, we observed a het-
designed for animal behavior studies, the Fish-CASU, erogeneous spatial repartition of the linear speed of the fish
Bonnet et al. 11

in the aquarium. Indeed, the fish tend to swim slowly in the acknowledge Daniel Burnier and Norbert Crot for their technical
corners and accelerate along the edges and even more so in support during the design and production of the Fish-CASU.
the center of the tank. Then, the detailed analysis of the
motion pattern showed that fish locomotion follows a Declaration of conflicting interests
three-step sequence: orientation, acceleration, and relaxa- The author(s) declared no potential conflicts of interest with respect
tion. First, the fish adjusts its direction. Then, it quickly to the research, authorship, and/or publication of this article.
accelerates in 2/15 s. Finally, it slowly decelerates to
return to its initial speed. The succession of these cycles Funding
produces a sinusoidal-like evolution of the acceleration The author(s) disclosed receipt of the following financial support
and linear speed. for the research, authorship, and/or publication of this article: This
We implemented a finite-state machine with three tun- work was supported by the EU-ICT project ASSISIbf, No.
able parameters into our FishBots to reproduce the zebra- 601074.
fish motion sequence underwater. Although there are small
differences, especially in terms of speed distribution, this References
is, to our knowledge, the first locomotion matching 1. Le Maho Y, Whittington JD, Hanuise N, et al. Rovers mini-
between a robotic device and zebrafish that shows that a mize human disturbance in research on wild animals. Nat
robot can behave similarly to a zebrafish. Moreover, the Methods 2014; 11: 1242–1244.
combination of the FishBot locomotion pattern and the 2. Halloy J, Sempo G, Caprari G, et al. Social integration of
body movements of the RiBot offers a wide range of sti- robots into groups of cockroaches to control self-organized
muli for behavior research. Indeed, by mimicking the choices. Science 2007; 318(5853): 1155–1158.
aspect ratio of the zebrafish and being able to reproduce 3. Gribovksiy A, Halloy J, Deneubourg JL, et al. The Poulbot, a
the tail beating frequency of the fish, the RiBot can emit mobile robot for ethological studies on domestic chickens. In:
different visual and hydrodynamic cues to interact with Proceedings of the international symposium on AI inspired
fish. The proposed controller also enables the precise and biology - A symposium at the AISB 2010 Convention, Leice-
constant adjustment of the acceleration and speed of the ster , United Kingdom of Great Britain and Northern Ireland,
Fish-CASU. While the shape and color of the fish lures are 29–1 April 2010, pp. 62–63.
important components of their attractiveness,6 the pattern 4. Gribovksiy A, Mondada F, Deneubourg JL, et al. Automated
motion of the robot could also play a key role in the com- analysis of behavioural variability and filial imprinting of
munication with groups of fish. Indeed, the fish perceive chicks (G. gallus), using Autonomous Robots. arXiv:1509.
and react to the movement of their congeners during col- 01957 2015.
lective motion. In particular, rapid changes in orientation or 5. Landgraf T, Oertel M, Rhiel D, et al. A biomimetic honeybee
movements performed in front of the group can propagate robot for the analysis of the honeybee dance communication
across the entire school, thanks to the network of visual system. In: The IEEE/RSJ 2010 international conference on
interaction of the fish.29–31 Therefore, the development of intelligent robots and systems (IROS 2010), Taipei, Taiwan,
a highly maneuverable robot that is able to reproduce fish 18–22 October 2010.
locomotion and quickly adjust its trajectory is an important 6. Abaid N, Bartolini T, Macri S, et al. Zebrafish responds dif-
step toward achieving artificial agents that can influence ferentially to a robotic fish of varying aspect ratio, tail beat
and lead collective motion by emitting similar visual and frequency, noise and color. Behav Brain Res 2012; 233:
kinetic signals to the fish. 545–553.
In future works, we will study the reaction of the fish to 7. Aureli M, Fiorilli F, and Porfiri M. Portraits of self-
the stimuli generated by the Fish-CASU. We will also organization in fish schools interacting with robots. Phys D
characterize the collective behaviors of mixed societies of 2012; 241(9): 908–920.
fish and robots and try to influence the collective choices of 8. Butail S, Polverino G, Phamduy P, et al. Fish-robot interac-
the fish society using our robotic platform. tions in a free-swimming environment: effects of speed and
configuration of robots on live fish. Bioinspir Biomim Proc.
of SPIE 2014, Vol. 9055: 90550I-1–90550I-7.
Author note
9. Phamduy P, Polverino G, Fuller RC, et al. Fish and robot
The information provided is the sole responsibility of the authors dancing together: bluefin killfish females respond differently
and does not reflect the European Commission’s opinion. The
to the courtship of a robot with varying color morphs. Bioin-
European Commission is not responsible for any use that might
spir Biomim 2014; 1–9.
be made of data appearing in this publication.
10. Marras S and Porfiri M. Fish and robots swimming together:
attraction towards the robot demands biomimetic locomotion.
Acknowledgements JR Soc Interface 2012; 9(73): 1856–1868.
The authors thank Marcelo Elias de Oliveira and Philippe 11. Faria J, Dyer J, Clément RO, et al. A novel method for inves-
Rétornaz for their assistance during the software and firmware tigating the collective behaviour of fish: introducing
implementation. Finally, the authors would like to gratefully “Robofish.” Behav Ecol Sociobiol 2010; 64: 1211–1218.
12 International Journal of Advanced Robotic Systems

12. Landgraf T, Nguyen H, Forgo S, et al. Interactive robotic fish 22. Zienkiewicz A, Barton DA, Porfiri M, et al. Data-driven sto-
for the analysis of swarm behavior. In: Tan Y, Shi Y and Mo chastic modelling of zebrafish locomotion. J Math Biol 2014;
H (eds) ICSI (1), ser. Lecture totes in computer science, Vol. 71(5): 1081–1105.
7928, Berlin: Springer, 2013, pp. 1–10. 23. Bonnet F, Rétornaz P, Halloy J, et al. Development of a mobile
13. Landgraf T, Bierbach D, Nguyen H, et al. RoboFish: robot to study the collective behavior of zebrafish. In: IEEE
increased acceptance of interactive robotic fish with realistic international conference on biomedical robotics and biome-
eyes and natural motion patterns by live Trinidadian guppies. chatronics (BioRob 2012), Rome, Italy, 24–27 June 2012.
Bioinspir Biomim 2016; 1–14. 24. Bonnet F, Binder S, Elias de Oliveira M, et al. A miniature
14. Swain DT, Couzin ID, and Leonard NE. Real-time feedback mobile robot developed to be socially integrated with species
controlled robotic fish for behavioral experiments with fish of small fish. In: IEEE international conference on robotics
schools. In: Proceedings of the IEEE, Vol. 100(1), 2012, pp. and biomimetics, Bali, Indonesia, 5–10 December 2014.
150–163. 25. Magnenat S, Rétornaz P, Bonani M, et al. ASEBA: a modular
15. Worm M, Landgraf T, Nguyen H, et al. Electro- architecture for event-based control of complex robots. In:
communicating dummy fish initiate group behavior in the IEEE/ASME transactions on mechatronics, Vol. 16(2),
weakly electric fish Mormyrus rume. Biomim Biohy Syst 2011, pp. 321–329.
2014; 8608: 446–448. 26. Magnenat S and Mondada F. ASEBA meets D-Bus: from
16. McGovern S, Alici G, Truong V, et al. Finding NEMO (novel the depths of a low-level event-based architecture into
electromaterial muscle oscillator): a polypyrrole powered the middleware realm. In: IEEE TC-Soft workshop on
robotic fish with real-time wireless speed and directional event-based systems in robotics (EBS-RO), St.Louis,
control. Smart Mater Struct 2009; 18(095009): 1–10. MO, USA, 10–15 october 2010.
17. Daou HE, Salumae T, Chambers LD, et al. Modelling of a 27. Merlin Systems Corp. “Merlin Robotics,” http://www.merlin
biologically inspired robotic fish driven by compliant parts. systemscorp.co.uk (accessed August 2016).
Bioinspir Biomim 2014; 9(1): 1–11. 28. Shi J and Tomasi C. Good features to track. In: Proceedings
18. Heo S, Wiguna T, Park HC, et al. Effect of an artificial caudal of the IEEE conference on computer vision and pattern
fin on the performance of a biomimetic fish robot propelled recognition, Seattle, USA, June 1994, pp. 593–600.
by piezoelectric actuators. J Bionic Eng 2007; 4(3): 151–158. 29. Lemasson BH, Anderson JJ, and Goodwin RA. Motion-guided
19. Wang Z, Hanga G, Li J, et al. A micro-robot fish with attention promotes adaptive communications during social
embedded SMA wire actuated flexible biomimetic. Sens navigation. Proc R Soc B 7 March 2013; 280: 20122003.
Actuat 2008; 144(2): 354–360. 30. Strandburg-Peshkin A, Twomey CR, Bode NW, et al. Visual
20. Nguyen QS, Heo S, Park HC, et al. Performance evaluation of sensory networks and effective information transfer in animal
an improved fish robot actuated by piezoceramic actuators. groups. Curr Biol 2013; 23: R709.
Smart Mater Struct 2014; 19(3): 1–8. 31. Rosenthal SB, Twomey CR, Hartnett AT, et al. Revealing the
21. Mwaffo V, Anderson RP, Butail S, et al. A jump persistent hidden networks of interaction in mobile animal groups
turning walker to model zebrafish locomotion. JR Soc Inter- allows prediction of complex behavioral contagion. Proc Natl
face 2015; 12(102): 20140844. Acad Sci USA 14 April 2015; 112: 4690–4695.