IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 7, NO.

4, OCTOBER 2022 9405

Event-Triggered Control of Robotic Fish With

Reduced Communication Rate
Wenyu Zuo, Animesh Chakravarthy , Member, IEEE, Michael Malisoff , Senior Member, IEEE,
and Zheng Chen , Senior Member, IEEE

Abstract—Underwater robots often need to communicate with robotic platforms for MUSN, bio-inspired underwater vehicles
external localization sensors. The low bandwidth in such commu- (BAUVs) have shown better maneuvering capabilities in shallow
nications is one of the bottlenecks in achieving accurate tracking water and have less impact on marine life [3], [4]. For example,
control. Toward this end, we adopt a novel periodic event-triggered compared to rotary propeller-driven autonomous underwater
control (PETC) which allows a robotic fish to reduce its commu- vehicles (AUVs) [5], [6], robotic fish use flapping tails for
nication load in tracking a desired heading angle with position
feedback from an external sensor. To design the PETC, a linear
propulsion, and this can eliminate acoustic noise caused by
state-space model is derived from a nonlinear dynamic model rotary propellers. BAUVs can achieve better stealth and higher
of the robotic fish with a small perturbation assumption. The propulsion efficiency than traditional AUVs [7]. Also, BAUVs
PETC consists of an observer, state-feedback controller, integrator, can swim smoothly in confined environments with grassy weeds
event-trigger rule, and predictor. The observer and state-feedback where traditional AUVs with rotational propellers cannot enter.
controller are designed to drive the tracking error to zero. The In addition to their noiseless operation, BAUVs can hover, hold,
integrator reduces the steady-state error. The event-trigger rule and move into restricted spaces.
determines when communication is needed while ensuring the effi- Although BAUVs have many advantages as MUSN, chal-
cacy of the state-feedback controller, and the predictor predicts the lenges arise when controlling BAUVs in GPS-denied under-
state vector for the state-feedback controller when communication water environments. Due to the limited onboard sensors for
is not available. For comparison, an observer-based state feedback
control (OSFC) and a proportional-integral-derivative (PID) con-
localization, BAUVs must rely on external sensors, such as
trol are implemented in real-time experiments. Simulations and a scanning sonar mounted on a mother-ship, to provide their
experimental results show that the PETC can dramatically re- location information remotely. The mother-ship needs to send
duce the number of communication instances without significantly location information to each BAUV through an underwater
degrading tracking performance, thereby saving communication acoustic network. Acoustic underwater communication has lim-
energy and reducing the need for high bandwidth underwater ited bandwidth, and its low data rate can lead to a reduced
communication. sampling rate for feedback control. Due to the under-actuated
Index Terms—Biologically-inspired robots, marine robotics,
dynamics of BAUVs, wireless control of BAUVs with such
motion control. drops in sampling rate has been a challenge. Hence, a control
method that is less communication-intensive is highly desirable
for controlling robotic fish.
I. INTRODUCTION To cover a large monitoring area with limited on-board power,
robotic fish usually cruise at their energy efficient speed and fol-
ESEARCHERS have developed different types of sensing
R networks to monitor distributed environmental parameters
in large coastal areas. Compared to fixed underwater sensing
low an optimal path with steering control. Since a robotic fish is
an under-actuated dynamic system whose aquatic working envi-
ronment contains nonlinearities and uncertainties, control design
networks, mobile underwater sensor networks (MUSN) have using hydrodynamic models is complicated. Existing steering
the advantages of longer duration, larger coverage, and recon- controls for robotic fish have mainly focused on error-based
figurable underwater monitoring [1], [2]. Among the existing proportional–integral–derivative (PID) control [8]–[10]. Some
nonlinear control algorithms have also been investigated, such
Manuscript received 29 January 2022; accepted 18 June 2022. Date of as collision-cone based guidance for collision avoidance [11],
publication 13 July 2022; date of current version 26 July 2022. This letter was sliding-mode control [12], and back-stepping control [13]. A
recommended for publication by Associate Editor Y. Peng and Editor X. Liu linear quadratic regulator yaw control was developed for a 3-
upon evaluation of the reviewers’ comments. The work of Michael Malisoff
was supported by NSF under Grant 2009659. (Corresponding author: Zheng
dimensional robotic fish [14]. However, experimental validation
Chen.) for this study was not provided.
Wenyu Zuo and Zheng Chen are with the Mechanical Engineering The above mentioned controls require control values to be
Department, University of Houston, Houston, TX 77204 USA (e-mail: computed at a fixed sampling frequency. This is not compu-
[email protected]; [email protected]). tationally efficient because it results in unnecessarily frequent
Animesh Chakravarthy is with the Mechanical and Aerospace Engineering
Department, University of Texas at Arlington, Arlington, TX 76019 USA computation and communication usage [15]. In many appli-
(e-mail: [email protected]). cations (especially networked systems), such inefficiencies in
Michael Malisoff is with the Department of Mathematics, Louisiana State computing controls have been addressed using event-triggered
University, Baton Rouge, LA 70803 USA (e-mail: [email protected]). control (ETC), which only recomputes control values when a
This letter has supplementary downloadable material available at
https://doi.org/10.1109/LRA.2022.3190612, provided by the authors.
significant event is detected [15]–[18]. Such events are usually
Digital Object Identifier 10.1109/LRA.2022.3190612 modeled as instances when a measurement from the system

2377-3766 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.
 9406 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 7, NO. 4, OCTOBER 2022

deviates from a predicted or reference value by more than robotic fish can mimic a real fish’s oscillating foil. The speed of
a prescribed amount. ETCs can have the significant benefit robotic fish depends on the caudal fin’s flapping frequency which
of saving bandwidth in communication networks by reducing is controlled by the DC motor’s speed. A servo motor outputs
the number of communication instances. Since ETC is less a bias angle to the robotic fish’s flapping so that the caudal fin
communication-intensive, it has great potential in swarming can generate a turning torque to steer the robot. In experiments,
control of robotic fish for MUSN. the fish is controlled by a micro-controller that translates the
This paper develops a model-based event-triggered control wireless command into a corresponding pulse width modulation
scheme for robotic fish to track a desired heading angle. We (PWM) signal. The control command is sent from a computer
consider a scenario in which the localization information is through WiFi. At the current stage of study, we only focus on
measured by external sensors, and control information is pro- the robotic fish’s two-dimensional control. Thus we only let the
vided by communication between the external sensor and the robotic fish swim at the water surface and manipulate WiFi to
robotic fish. A linear steering state-space model is derived by simulate underwater communication.
simplifying a full nonlinear hydrodynamic model [19], [20]
following a perturbation-based Nomoto model, which is a well- B. Nonlinear Dynamic Model
developed perturbation model for vessel control [21]. Using
this linear model, we adopt a periodic event-triggered control The linear model for PETC is obtained from a full nonlinear
(PETC) [22] that uses a predictor to estimate the state vector in dynamic model through a linearization process. The nonlinear
open-loop. The PETC is derived in the discrete-time domain dynamic model of the robotic fish is established in 2-dimensional
from an observer-based state feedback control (OSFC) [23]. Cartesian coordinates [26]. We use O − XY as the earth-fixed
The PETC is first analyzed using simulations to examine its coordinate, and G − ij as the body-fixed coordinate. The robot’s
effectiveness in steering the fish in desired directions, and the mass center is coincident with the body-fixed origin G. In the
result shows that the PETC can achieve similar performance to body-fixed coordinate, u is the surge in the i-direction, v is
a non-PETC controller but with a significantly reduced number the sway in the j-direction, and r is the angular velocity. The
of communication instances. The PETC also exhibits better velocities are transferred from body coordinates to earth-fixed
tracking performance when encountering communication rate coordinates using the equations
drops. We also provide experimental results that were obtained Ẋ = u cos ψ − v sin ψ,
using our prototype robotic fish, and we believe that this pa-
per shows the first experimental implementation of PETC on Ẏ = u sin ψ + v cos ψ,
robotic fish.
The major contributions of this paper are as follows. We ψ̇ = r, (1)
address the real-world challenge of how to wirelessly control where ψ is the yaw angle.
robotic fish when the communication bandwidth and power In the robot’s body-fixed coordinates, the thrust, damping
are limited. An individual robot may encounter a communi- force, and body added-mass can be described by [27], [28]:
cation rate drop due to a large number of messages flooding
the communication network. We use the event-trigger rule to (m − Xu̇ ) u̇ − (m − Yv̇ ) vr + X|u| u = Ti ,
reduce the need for state vector transmission by communication
between the external sensor and the robotic fish. We also study (m − Yv̇ ) v̇ + (m − Xu̇ ) ur + Y|v| v = Tj ,
the effectiveness of PETC in controlling robotic fish and in (Iz − Nṙ ) ṙ − (Yv̇ − Xu̇ ) uv + N|r| r = NT , (2)
estimating a state vector in open-loop. PETC reduces the use
of communication and helps the robot maintain stability even where m is the mass of robotic fish, Iz is the moment of inertia.
when the communication rate is reduced. Our PETC design X|u| , Y|v| , and N|r| are the linearized drag coefficients, and Xu̇ ,
is based on a linear model that is obtained by linearizing the Yv̇ , and Nṙ are the constant hydrodynamic added-mass coeffi-
nonlinear dynamic model of robotic fish about a straight swim- cient, calculated assuming that the fish body has an ellipsoidal
ming equilibrium condition. The PETC has a predictor that uses shape. Their values are obtained through
model information to estimate the state vector without knowing Xu̇ = −k1 m, Yv̇ = −k2 m, Nṙ = −k  Iz , (3)
the output. To ensure tracking accuracy, we add integrators to
reduce the steady-state error [24]. Simulations and experiments where k1 , k2 , and k  are added mass coefficients from [29]. Also,
demonstrate the effectiveness of the proposed PETC design. Ti and Tj are the components of thrust acting along the i and
The rest of the paper is organized as follows. A prototype j axes, respectively. They are calculated in an additional DSC
of the robotic fish and its dynamic model are described in coordinate using a vectorized EBT method [20]. The calculation
Section II. The ETC is described in Section III. Simulation only considers forces produced by the caudal fin’s flapping,
and experimental results are presented in Section IV. Finally, and its output forces are Fid and Fjd , which are force vectors
conclusions and future work are described in Section V. along the DSC coordinate’s i and j axes, respectively. Hence,
the component forces Ti and Tj are
    
II. ROBOTIC FISH AND ITS MODELING Ti cos δ − sin δ Fid
= kF , (4)
Tj sin δ cos δ Fjd
A. Robotic Fish
where kF is a coefficient which can be tuned by model fitting.
In this study, the robotic fish has a 3-joint tail which includes Also, NT denotes the net turning moment caused by Ti and Tj ,
a 6V servo motor and a double-slider-crank (DSC) mecha- that is,
nism [25]. The DSC mechanism uses one DC motor to achieve
a two-degree-of-freedom flapping of the caudal fin so that the NT = Tj (d0 + l cos δ) − Ti l sin δ, (5)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.
 ZUO et al.: EVENT-TRIGGERED CONTROL OF ROBOTIC FISH WITH REDUCED COMMUNICATION RATE 9407

TABLE I obtained through a hydrodynamic simulation from our previous

MODEL PARAMETERS work [25]. Then the state-space representation of the sway and
yaw motion is
   Y|v|    ⎡ kF Fi ⎤
Δv̇ − m1 − mm2 u1 0 Δv d

= + ⎣ m1 ⎦ Δδ, (10)
N
Δṙ 0 − m|r|3 Δr kF d0 Fid
m3

where we use the notation

m1 = m − Yv̇ , m2 = m − Xu̇ , and m3 = Iz − Nṙ . (11)
Following [25], we convert this state space equation with C =
[0, 1] into a transfer function that describes the yaw rate of a
robotic fish corresponding to a bias angle input, to obtain
Δr(s) a1 s + a2
= 2 , (12)
Δδ(s) s + a3 s + a4
Fig. 1. A schematic and a photo of the robotic fish.
where
where δ is the angle of the first joint which is equivalent to a kF d0 Fid Y|v| kF d0 Fid
bias angle of the oscillating foil, d0 is the distance from the a1 = , a2 = ,
m3 m1 m3
fish’s center of mass to the first joint, and l is the distance from m3 Y|v| + m1 N|r| Y|v| N|r|
the first joint to the caudal fin’s mass center. Table I shows the a3 = , and a4 = .
parameters of the model obtained from [25]. See also Fig. 1 for m1 m3 m1 m3
a schematic of and a photo of the robotic fish. The transfer function has two negative poles given by
Y N
− m|v|1 , − m|r|3 . When a robotic fish is swimming, the measure-
C. Linear Steering Dynamic Model
ment of yaw rate is heavily disturbed by the lateral oscillations
The linear steering model is derived from a Namoto model of the fish head. Adding an integrator to (12) to derive a transfer
which captures a vessel’s yaw dynamics. The robotic fish is function between the bias angle and the yaw angle is more
assumed to swim with a constant flapping frequency. Following feasible for implementation purposes. With (1), (6), and (12),
the notation from [26], its velocities (u, v, r) are we obtain
u = u0 + Δu, v = Δv, and r = Δr, (6) ψ(s) 1 Δr(s)
= · . (13)
where u0 is the constant initial forward speed, (Δu, Δv, Δr) Δδ(s) s Δδ(s)
are the velocity perturbations, and the time derivatives of the Because the sensing and communication of the robotic fish
velocities are are running on a digital device, the controller is ideally to be
u̇ = Δu̇, v̇ = Δv̇, and ṙ = Δṙ, (7) designed in the discrete-time domain. A discrete state-space
representation of (13) with a 2Hz sampling frequency is obtained
respectively. The bias angle δ in forward swimming is also as follows
considered as a perturbation, and so is denoted as Δδ. For a
x∗ (k + 1) = Ax∗ (k) + Bu∗ (k),
small bias angle satisfying |Δδ| ∈ (0, 0.35], one can take the first
order Taylor expansion to make the approximations cos Δδ ≈ 1 y ∗ (k) = Cx∗ (k), (14)
and sin Δδ ≈ Δδ. Since a small bias angle has minimal effect
on the thrust magnitude along the i-axis, one can assume that where, x∗ (k) ∈ R3×1 is the state vector, u∗ (k) ∈ R is the input
bias angle, and y ∗ (k) ∈ R is the output yaw angle ψ. In this
Ti = Ti0 , Tj = ΔTj , and NT = ΔNT , (8) study, the system matrices A, B, and C are considered time-
are satisfied with a constant Ti0 and the initial values Tj0 = invariant when the robotic fish is swimming in calm water at a
NT 0 = 0 for Tj and NT . Then (2) can be rewritten as constant speed and being controlled by a small bias angle. Hence
the value of A, B and C are obtained from our earlier work [25]
(m − Xu̇ )Δu̇ + X|u| u0 + X|u| Δu = Ti0 , as follows:
⎡ ⎤ ⎡ ⎤
(m − Yv̇ )Δv̇ + (m − Xu̇ )u0 Δr + Y|v| Δv = ΔTj , 2.832 −1.335 0.8381 0.25
A=⎣ 2 0 0 ⎦, B = ⎣ 0 ⎦, (15)
(Iz − Nṙ )Δṙ + N|r| Δr = ΔNT , (9)
0 0.5 0 0
by ignoring the high order perturbations, and the Coriolis angular
momentum Yv̇ − Xu̇ caused by the added mass is ignored to and C = −0.1232 0.1785 0 . One can conclude that the
moderate the yaw oscillation caused by the recoil head move- system with (A, B, C) is fully controllable and observable.
ment. The mean thrust along the i axis during a full flapping We conducted experiments to validate the linearized model.
cycle is denoted by Fid . Also, the mean thrust Fjd = 0 because Fig. 2 shows the results of model validation, where δ0 represents
the mean value of the lateral force over a full flapping cycle the neutral position of tail flapping. The model prediction and
is zero. Since the robotic fish is assumed to be swimming at experimental data are compared at each bias angle. The average
a constant flapping frequency, the value Fid = −0.035N is prediction error after 8 seconds is 5%.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.
 9408 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 7, NO. 4, OCTOBER 2022

Fig. 3. ETC Diagrm.

steering. The error dynamics model is obtained from the state-

Fig. 2. Results of model validation. space model in (14):
x(k + 1) = Ax(k) + Bu(k),
III. CONTROL SYSTEM DESIGN y(k) = Cx(k), (18)
There are three challenges in designing a linear controller for where x, u, and y are defined by
the robotic fish steering. First, underwater communication has
limited bandwidth, which reduces the sampling rate available x = x∗ − x∗ss , u = u∗ − u∗ss , and y = y ∗ − ψd ,
for performing closed-loop feedback control. To address this and where A, B, and C are from Section II-C. x∗ss is the steady-
issue, we propose to use PETC to reduce the need for a high state value of x∗ , u∗ss is the steady-state value of u∗ , and ψd is a
sampling rate. Second, a small underwater disturbance can cause non-zero desired yaw angle.
a steady-state error in heading angle that leads to a large position An integrator that integrates the heading angle tracking error
error in a long mission. To address this issue, we introduce an over time is defined as
integrator into the ETC, to reduce the steady-state error. Third,
to ensure the linearized model’s accuracy for PETC, a control q(k + 1) = q(k) + Cx(k). (19)
constraint is added to the controller. To avoid control saturation, Hence the full dynamics of the robotic fish with the integrator
we introduce a task divider to divide a large turning angle task can be written as
into a series of small turning angle subtasks.       
x(k + 1) A 0 x(k) B
= + u(k), (20)
q(k + 1) C I q(k) 0
A. Event-Triggered Control    
ETC considers a fully controllable discrete dynamic system A B
that is controlled by a state-feedback controller. However, the where (A, B  ) is controllable. The control law is chosen as
state vector is not sent to the controller at every time instant.  
They are only sent when an ‘event’ (where the rule for defining x(k)
u(k) = −K H , (21)
the event is specified below) occurs, and the ‘event’ triggers the    q(k)
sending of the state vector. We assume that the state vector is K
sent at time k = 0. At any integer time k > 0, we define k − to
be the previous event-triggered time, and the control input is ob- with a state feedback gain K ∈ R1×3 and an integral gain H ∈
tained through a zero-order-hold when k = k − . Thus, the control R. Define x (k) = [x(k), q(k)]T . Then the closed-loop system
law is becomes
 
A − BK BH 
u(k) = K  x (k − ). (16) x (k + 1) = x (k). (22)
C I
  
with a state feedback gain K  ∈ R1×3 and a fully controllable Λ
state vector x (k) ∈ R3×1 . Let e(k) denote the error e(k) =  
Since (A , B ) is controllable, a pole assignment approach can
x (k − ) − x (k) between the value of the held state and the
be used to select the gain K  such that Λ is a Hurwitz matrix in
real-time state. The event-trigger rule is defined as follows: For
discrete time. Then (22) implies that x (k) converges zero and
a given constant σ > 0, an event is triggered when the inequality
limk→∞ y(k) = 0.
A Lyapunov function is built as V (x ) = xT P x , where P is a
e(k) ≤ σ x (k) (17)
suitable symmetric positive definite matrix. Then, the difference
δV (x (k)) = V (x (k + 1)) − V (x (k)) satisfies
is violated. Hence, (17) holds for all k ≥ 0. The constant σ ≥ 0
will be specified in the control design, and  ·  is the usual δV (x (k)) = xT (k)(ΛT P Λ − P )x (k) < 0, (23)
Euclidean norm (as used in [30]). A diagram of the ETC scheme
is shown as Fig. 3 below. for all nonzero states x (k) of (21), because ΛT P Λ − P is
negative definite.
If we redesign the controller following (16), then (20) is
B. Stability Analysis With ETC rewritten as
The steering of robotic fish is controlled by a state-feedback x (k + 1) = A x (k) + B  K  (e(k) + x (k)),
(24)
control law based on the error dynamics model of robotic fish = Λx (k) + B  K  e(k).

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.
 ZUO et al.: EVENT-TRIGGERED CONTROL OF ROBOTIC FISH WITH REDUCED COMMUNICATION RATE 9409

heading angle tracking error. When the event is triggered, the

communication between the robot and the remote controller is
established. The variables z1 and z2 in Predictor-1 and Predictor-
2, respectively, will be replaced by the estimated state vector
x, and the variables q1 and q2 in Integrator-1 and Integrator-2,
respectively, will be replaced by q3 in Integrator-3.
In the remote controller, a sensor measures the robotic fish’s
yaw angle ψ and provides y to the observer. The observer is
defined as
x̄(k + 1) = Ax̄(k) + Bu(k) + L[y(k) − C x̄(k)], (27)
Fig. 4. Control diagram of PETC. 3×1 3×1
where x̄(k) ∈ R is the estimate of x(k) and L ∈ R is the
observer gain that ensures the convergence of the observer.
In this ETC case, δV satisfies In the ‘non-triggered’ mode, we follow the embedded pre-
dictor idea that was proposed by [22], to have an open-loop
δV (x (k)) = −xT (k)Qx (k) predicted state vector z. The use of z can alleviate the need for
+2xT (k)ΛT P B  K  e(k) (25) a zero-order hold of the control input u when x̄ is unavailable
+ eT (k)K T B T P B  K  e(k), to the robot. The dynamics of the predictor are defined by
where Q = −ΛT P Λ + P is a positive definite matrix. 
Azi (k) + Bu(k) (Non − triggered),
Combining (25) and (17) gives zi (k + 1) = (28)
Ax̄(k) + Bu(k) (Triggered),
δV (x (k)) ≤ −xT (k)Qx(k)  where zi ∈ R3×1 , and i = 1, 2. Then z1 represents Predictor-
+ 2 x (k) ΛT P B K   e(k)
K T B T P B  K   e(k)2 1 whose dynamics are defined in (28). Predictor-2 (which is
+   (26) defined by z2 ) has the same dynamics as z1 . These two predictors
≤ − λmin [Q] − 2σ ΛT PB  K   run simultaneously and work at the same sampling rate as the
−σ 2 K T B T P B  K   x (k)2 , remote controller. These two predictors use the same system
matrices (A, B) as the ones used in the observer to estimate the
where λmin [Q] is the smallest eigenvalue of Q. There is a value state vector.
ε > 0 for each σ ∈ (0, ε], δV (x (k)) < 0 when x (k) = 0, so Through integration of the output error, Integrator-3 is used to
the system achieves an asymptotic stability property. reject the influence of external disturbances such as water waves.
The external disturbance is then actively rejected by adding an
C. PETC Implementation extra term in the control u. Integrator-1, represented by q1 ∈ R,
The traditional ETC uses the last stored state vector to calcu- runs inside the robot. Integrator-2 q2 ∈ R runs in the remote
late the control input, which puts the controller in zero-order- controller. Then q1 and q2 have the dynamics
hold, and is not ideal for steering control. PETC has predictors 
q (k) + Czi (k) (Non − triggered),
that can estimate the state vector in an open loop [22]. It can qi (k + 1) = i (29)
prevent zero-order hold when the ‘event’ is not triggered. The q3 (k + 1) (Triggered),
mechanism of PETC is introduced as follows. for i = 1, 2. Integrator-3 is running in the remote controller and
In PETC, the feedback information is measured by a remote has the dynamics
controller, and the state vector is sent from the remote controller
to robotic fish using communication. There are two modes of q3 (k + 1) = q3 (k) + y(k). (30)
operations, namely, ‘triggered’ mode and ‘non-triggered’ mode. It is updated using y to represent the ‘true’ error, similar to the
In the ‘triggered’ mode, communication is available between the observer representing the ‘true’ state vector. With x̄, z and q as
robotic fish and the remote controller. The control law inside the defined, the control input u discussed in (16) is replaced by
robotic fish works as a traditional state feedback control using 
the state vector sent from the remote controller. In the ‘non- −Kz1 (k) + Hq1 (k) (Non − triggered),
triggered’ mode, communication between the remote controller u(k) = (31)
−K x̄(k) + Hq3 (k) (Triggered).
and the robot is unavailable.
The overall control design can be described by the diagram Here, H ∈ R is an integral gain. Also, u is calculated using
shown in Fig. 4, where the green block represents the dynamics the ‘true’ error and the ‘true’ state vector from the remote
of the robot with its local controller and the orange block controller in the ‘triggered’ mode. In the ‘non-triggered’ mode,
represents the dynamics of the remote controller. In the robot, u is calculated using the error and state vector estimated inside
Predictor-1 is used to estimate the state vector when the com- the robot. The quantities z2 and q2 agree with the calculation of
munication is not triggered. Integrator-1 is used to integrate the u for the observer because they have the same dynamics as z1
predicted heading angle tracking error. The control u combines and q1 .
the state feedback control and integration of the tracking error, The event-trigger rule is interpreted as comparing the state
with different gains K and H. In the remote controller side, a vector error between the predictor and the observer.
computer vision sensor is used to calculate the heading angle 
z2 (k) − x̄(k) ≤ σ x̄(k) ⇒ (Non − triggered),
y. Then, y is sent to the observer to estimate the state vector
x. Predictor-2 is the same as Predictor-1. Integrator-2 is the z2 (k) − x̄(k) > σ x̄(k) ⇒ (Triggered).
same as Integrator-1. Integrator-3 is used to integrate the actual (32)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.
 9410 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 7, NO. 4, OCTOBER 2022

Fig. 5. Task divider.

A constant σ ≥ 0 is to be specified as a threshold. The event- Fig. 6. Simulation result of PETC for a step reference.
trigger rule is running concurrently with the state dynamics and
determines which mode the system operates in at each time
k > 1. The condition
z2 (k) − x̄(k) > σ x̄(k) (33)
is defined as the ‘event,’ indicating that the predictor’s estimation
accuracy is unsatisfied.
The advantages of adding the predictors can be summarized
as follows. First, the control input u applied to the plant is
updated even during the non-triggered mode (using (31)) and
this alleviates the zero-order hold that would have been used
otherwise. Second, the control input u supplied to the observer
in the remote controller (whose formula is given by (31) with
z2 and q1 replaced by z2 and q2 , respectively), can be calculated Fig. 7. Simulation result of PETC with different σ.
using z2 (k) in the non-triggered mode, thereby indicating that
communication is not necessary at every instance. see that the linear model can still predict the yaw angle with
A task divider is added to divide a complex maneuvering the bias angle equal to 0.5 rd. A constant communication rate
task into multiple sub-tasks such that each sub-task runs in an (CCR) of 2Hz is the default sampling rate. The controller design
individual sub-frame. As shown in Fig. 5, the purpose of the is based on the linear model (18) and observer (27), with
task divider is to break down a large amplitude reference step
input into piecewise-constant step inputs of smaller amplitudes. L = [−36.23, −76.05, −40.04]T ,
The aim is to prevent the control input from getting saturated. K = [3.45, −3.04, 2.67], and H = 0.05 (34)
Given a constant umax > 0, the objective is to ensure that ψd
is sufficiently small so that u(k) in the feedback control is such as the coefficient matrices.
that |u(k)| < umax for all k.
B. Simulation Results
IV. SIMULATION AND EXPERIMENTAL RESULTS In Fig. 6(a), PETC with σ = 0.2 to track the ψd = 0.4 rd
is shown. Its bias control input u(k) is plotted as a solid line
A. Simulation Configuration
in Fig. 6(b). We view the reception of x̄(k) as a triggered
In the simulation, the nonlinear fish dynamics is simulated communication from the remote controller to the robot, and use
using a 2nd-order Runge-Kutta method with (1) and (2). The a red circle mark to represent every such trigger. The PETC
simulation assumes that the robotic fish’s position is obtained had a total of 16 triggers during the simulation. The ‘event’
through an external sensor at a constant sampling rate. The yaw plot Fig. 6(c) shows how the event is triggered in terms of
angle ψ(k) is calculated using a position data sequence. The (32) in Section III-C. When the trigger is applied, the value
position acquisition in real circumstances may contain sensing of z2 (k) − x̄(k) is immediately lowered. The predicted state
errors, and the yaw angle calculation is highly disturbed by the z2 and observed state x̄(k) are shown in Fig. 6(d).
fish head’s swing movement. A Gaussian noise of ± 0.02 m Fig. 7 shows how PETC behaves with various values of σ. It
magnitude is added to the obtained fish position. The control is clear from Fig. 7(a) that σ = 0.1 causes the most triggers.
objective is to control the robotic fish’s swimming direction to A smaller σ leads to not only a lower threshold, but also more
a desired reference yaw angle ψd . In simulations, the robotic frequent communication to constrain the z2 (k) − x̄(k). The
fish starts from a standstill and swims with a constant flapping σ value can be selected based on how much communication
frequency of 2 Hz. After 3 seconds of straight swimming, the resource needs to be saved.
controller starts to steer the robot. In the experiments, the small While PETC can reduce the number of triggers and con-
bias angle limits the maneuverability of the robotic fish. Thus, serve communication resources, its performance can also be
in the implementation, we increase the input boundary to 0.4 rd maintained when the communication sampling rate drops. The
and consider the error caused by this input increment to be controller was tested in a reduced communication rate (RCR)
unmodeled dynamic. We also added a plot showing the model case, in which the communication sampling rate drops to 0.3Hz.
prediction versus the actual yaw angle output in Fig. 2. We can An OSFC is designed for comparison. When the communication

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.
 ZUO et al.: EVENT-TRIGGERED CONTROL OF ROBOTIC FISH WITH REDUCED COMMUNICATION RATE 9411

Fig. 8. OSFC and PETC at 0.3 Hz communication rate.

Fig. 10. Experimental Trajectories.

In Fig. 8(b), PETC in the RCR case is shown. A much smaller

steady-state error in ψ is exhibited, and u(k) has only 5 triggers
with no obvious oscillation.

C. Experimental Results
In experiments, the robot swam in an above-ground swim-
ming pool that was 10 meters long and 5 meters wide. The
observer and event-trigger rule ran on a laptop computer which
communicated with the robotic fish through WiFi. The robotic
fish’s position was obtained through computer vision using a
camera installed 5m above the pool. The sensing noise arises
from the error in computer vision and waves in the swimming
pool. The feedback gain was adjusted to K=[3.7, -3.33, 2.93]
through experiment tuning. The control task is a two-step track-
ing: The first step tracking requires the robotic fish to steer to
0.25π rad in 15 seconds with a 2Hz communication rate. The
second step tracking requires the robotic fish to steer to 0.5π rad
with a 0.3Hz communication rate. Besides PETC and OSFC, we
implemented a PID controller with Kp = 0.6, Ki = 0.005, and
Kd = 0.4 for comparison.
In Fig. 9(a), the result of PETC with σ = 0.2 is shown.
We use a red circle to represent a reception of x̄(k). In the
first step tracking, PETC has 8 triggers in total, and its output
Fig. 9. Experimental results. has a negligible overshoot with an average 9 s settling time.
The ‘event’ subplot shows that the event-trigger rule makes
is unavailable, the situation is equivalent to a non-triggered the prediction error converge to zero in the CCR case. In the
mode. Hence the control law for OSFC can be represented as second step tracking, the event-trigger rule has less capabil-
 ity to constrain the prediction error. Hence the output shows
−K x̄(k − ) + Hq3 (k − ) (Non − triggered), more overshoot and oscillation. An experimental OSFC result
u(k) = (35) is shown in Fig. 9(b). The controller received x̄(k) from the
−K x̄(k) + Hq3 (k) (Triggered).
observer at every communication instance after 5 seconds of
using the same definition of k − that we used in (16). Fig. 8(a) straight swimming. In the CCR case, OSFC has a good tracking
shows how OSFC behaves in the RCR case. Because only 11 result with 30 triggers. After that, the communication rate drops
triggers are available and u(k) is held constant between triggers, to 0.3Hz, OSFC only has 5 triggers, and the u(k) is held constant
OSFC has a large error in ψ and the u(k) exhibits oscillations. between two triggers. Consequently, there is a more obvious

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.
 9412 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 7, NO. 4, OCTOBER 2022

TABLE II [5] X. Tan, “Autonomous robotic fish as mobile sensor platforms: Challenges
CONTROLLER PERFORMANCE and potential solutions,” Mar. Technol. Soc. J., vol. 45, no. 4, pp. 31–40,
2011.
[6] R. Du, Z. Li, K. Youcef-Toumi, and P. V. Y Alvarado, Robot Fish: Bio-
Inspired Fishlike Underwater Robots. Berlin, Germany: Springer, 2015.
[7] M. Sfakiotakis, D. M. Lane, and J. B. C. Davies, “Review of fish swimming
modes for aquatic locomotion,” IEEE J. Ocean. Eng., vol. 24, no. 2,
pp. 237–252, Apr. 1999.
[8] P. Phamduy, J. Cheong, and M. Porfiri, “An autonomous charging system
for a robotic fish,” IEEE/ASME Trans. Mechatronics, vol. 21, no. 6,
oscillation in the yaw plot after 20 seconds. The PID control pp. 2953–2963, Dec. 2016.
[9] Z. Su, J. Yu, M. Tan, and J. Zhang, “Closed-loop precise turning control
exhibits a similar result in Fig. 9(c). for a BCF-mode robotic fish,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots
The trajectories for the above three experiments are plotted Syst., 2010, pp. 946–951.
in Fig. 10(a). A trajectory from PETC with H = 0 is added [10] C. Meurer, A. Simha, Ü. Kotta, and M. Kruusmaa, “Nonlinear orientation
to demonstrate the integrator. The robotic fish starts from an controller for a compliant robotic fish based on asymmetric actuation,” in
initial yaw angle of 0 degrees, then finishes with two 45-degree Proc. Int. Conf. Robot. Automat., 2019, pp. 4688–4694.
[11] X. Yi, A. Chakravarthy, and Z. Chen, “Cooperative collision avoidance
step tracking. The yaw oscillation of the OSFC and PID in the control of servo/IPMC driven robotic fish with back-relaxation effect,”
RCR case are shown as green and blue lines, and the PETC IEEE Robot. Automat. Lett., vol. 6, no. 2, pp. 1816–1823, Apr. 2021.
trajectory (with H = 0) has a more obvious steady-state error. [12] H. Wang et al., “Precise discrete-time steering control for robotic fish based
Among them, the trajectory of PETC is shown in Fig. 10(b) to on data-assisted technique and super-twisting-like algorithm,” IEEE Trans.
Ind. Electron., vol. 67, no. 12, pp. 10587–10599, Dec. 2020.
exhibit that the PETC can stably control the robotic fish while it [13] M. L. Castaño and X. Tan, “Backstepping control-based trajectory tracking
is swimming forward and steering. The control performances in for tail-actuated robotic fish,” in Proc. IEEE/ASME Int. Conf. Adv. Intell.
terms numbers of triggers and mean square error at steady-state Mechatronics, 2019, pp. 839–844.
are shown in Table II. Each control is tested in 5 repeated [14] P. Suebsaiprom and C. Lin, “Maneuverability modeling and trajectory
experiments to collect the data. Table II shows that the PETC sac- tracking for fish robot,” Control Eng. Pract., vol. 45, pp. 22–36, 2015.
[15] W. Heemels, K. Johansson, and P. Tabuada, “An introduction to event-
rifices some tracking accuracy to save communication resources, triggered and self-triggered control,” in Proc. 51st IEEE Conf. Decis.
and its performance is consistent when the communication rate Control, 2012, pp. 3270–3285.
drops. [16] A. Selivanov and E. Fridman, “Distributed event-triggered control
of transport-reaction systems,” IFAC-PapersOnLine, vol. 48, no. 11,
pp. 593–597, 2015.
V. CONCLUSION AND FUTURE WORK [17] A. Selivanov and E. Fridman, “Event-triggered H∞ control: A switching
approach,” IEEE Trans. Autom. Control, vol. 61, no. 10, pp. 3221–3226,
We developed PETC for robotic fish to reduce the commu- Oct. 2016.
nication burden between the robot and a remote sensor. The [18] A. Selivanov, E. Fridman, and A. Fradkov, “Event-triggered adaptive
PETC was compared with OSFC and PID in simulations and control of minimum-phase systems,” IFAC-PapersOnLine, vol. 50, no. 1,
experiments. Our results show that the PETC with appropriate pp. 4276–4281, 2017.
trigger thresholds can reduce the number of communication [19] M. J. Lighthill, “Note on the swimming of slender fish,” J. Fluid Mechan-
ics, vol. 9, no. 2, pp. 305–317, 1960.
instances while still achieving performance that is as good as [20] J. Wang and X. Tan, “A dynamic model for tail-actuated robotic fish with
OSFC. Our results also illustrate that with a reduced commu- drag coefficient adaptation,” Mechatronics, vol. 23, no. 6, pp. 659–668,
nication rate, PETC can maintain performance in steering and 2013.
straight swimming. Our use of an integrator and a task divider [21] K. Nomoto, T. Taguchi, K. Honda, and S. Hirano, “On the steering qualities
of ships,” Int. Shipbuilding Prog., vol. 4, no. 35, pp. 354–370, 1957.
provides significant novelty, compared to prior treatments of [22] W. Heemels and M. Donkers, “Model-based periodic event-triggered
PETC. control for linear systems,” Automatica, vol. 49, no. 3, pp. 698–711, 2013.
Since the controller’s performance relies on the accuracy of [23] O. Katsuhiko, Modern Control Engineering, 5th ed. Harlow, U.K.: Pearson
the model, internal model uncertainty could introduce a state and Edu. Limited, 2010.
input dependent disturbance [31] which is difficult to reject in [24] G. F. Franklin, J. D. Powell, A. Emami-Naeini, and J. D. Powell, Feedback
Control Dynamic Syst., vol. 4. Hoboken, NJ, USA: Prentice Hall, 2002.
PETC. Our future work will focus on developing an equivalent- [25] W. Zuo, F. E. Fish, and Z. Chen, “Bio-inspired design, modeling, and con-
input-disturbance (EID) estimator which can estimate the EID trol of robotic fish propelled by a double-slider-crank mechanism driven
and directly reject it in PETC. In addition, the benefits of tail,” J. Dynamic Syst. Measur. Control, no. 12, 2021, Art. no. 121005.
using PETC in multi-agent and multi-task scenarios will also [26] T. I. Fossen, “Guidance and control of ocean vehicles,” Ph.D. thesis, Univ.
Trondheim, Trondheim, Norway, 1999.
be explored. [27] Z. Chen, P. Hou, and Z. Ye, “Robotic fish propelled by a servo motor
and ionic polymer-metal composite hybrid tail,” J. Dynamic Syst. Measur.
REFERENCES Control, vol. 141, no. 7, pp. 1–11, 2019.
[28] J. Wang and X. Tan, “Averaging tail-actuated robotic fish dynamics
[1] Y. Ryuh, G. Yang, J. Liu, and H. Hu, “A school of robotic fish for through force and moment scaling,” IEEE Trans. Robot., vol. 31, no. 4,
mariculture monitoring in the sea coast,” J. Bionic Eng., vol. 12, no. 1, pp. 906–917, Aug. 2015.
pp. 37–46, 2015. [29] H. Ghassemi and E. Yari, “The added mass coefficient computation of
[2] Y. Wang et al., “Aquatic debris monitoring using smartphone-based robotic sphere, ellipsoid and marine propellers using boundary element method,”
sensors,” in Proc. 13th Int. Symp. Inf. Process. Sensor Netw., 2014, pp. 13– Polish Maritime Res., vol. 18, no. 1, pp. 17–26, 2011.
24. [30] H. K. Khalil, Nonlinear Systems, 3rd ed. Upper Saddle River, NJ, USA:
[3] F. E. Fish, “Advantages of aquatic animals as models for bio-inspired Prentice Hall, 2002.
drones over present AUV technology,” Bioinspiration Biomimetics, [31] R. Liu, G. Liu, M. Wu, F. Xiao, and J. She, “Robust disturbance rejection
vol. 15, no. 2, 2020, Art. no. 025001. based on equivalent-input-disturbance approach,” IET Control Theory
[4] F. E. Fish, “Bio-inspired aquatic drones: Overview,” Bioinspiration Appl., vol. 7, no. 9, pp. 1261–1268, 2013.
Biomimetics, vol. 15, no. 6, 2020, Art. no. 060401.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on January 16,2025 at 15:50:28 UTC from IEEE Xplore. Restrictions apply.