12th IEEE Workshop on Variable Structure Systems,
VSS’12, January 12-14, Mumbai, 2012
Sliding Mode Control Design for a Carangiform Robotic Fish
Jian-Xin Xu, Xue-Lei Niu and Zhao-Qin Guo
Abstract— In this paper, measures to resist parameter un-
certainties and external disturbances are presented for a
biomimetic robotic fish. The Carangiform robotic fish consists
of 𝑁 links and 𝑁 − 1 joints, and its dynamic model of motion
is given in previous work [1]. Through this model, the relation
between the motion of the fish and the torques added are
constructed. By giving particular reference angles of joints,
forward locomotion is obtained and serves as an example for
the following study. Aiming at parameter uncertainties and
environmental disturbances, we adopt sliding mode control
(SMC) to attenuate their impacts. Results of numerical ex-
amples validate the effectiveness of SMC. Comparison of those
results shows that SMC works better than traditional computed
torque control when performing the tracking task in joint space.
I. I NTRODUCTION
With increasing underwater activities and research work,
such as underwater archaeology, leakage detection, military
reconnaissance [2], Autonomous Underwater Vehicle (AUV)
is receiving more and more attention [3]. Traditional AUV,
usually thrusted by rotatory propellers, may not be satis-
factory in efficiency and maneuverability. Thus, new type
of AUV is needed. During the long period time of nature
selection, fish have evolved body structures and swimming
patterns that highly adapt to aquatic environments [4]. Ac-
tually, they are more advanced swimming machines with
higher efficiency, more remarkable maneuverability and less
noise than conventional AUV.
Inspired by these appealing merits, researchers developed
many theories and numerous robotic fish prototypes to
study and mimic the way that real fish moves. Generally,
swimming modes of fish is classified to three main cate-
gories: Anguilliform, Carangiform and Thunniform [5], [6].
In Anguilliform mode, the whole body participates in large
amplitude undulations. While in Carangiform swimming, the
fish body undulations are confined to the last third of the
body length, and in Thunniform, the undulation proportion
is even less. Elongated body theory (EBT) [7], [8], assuming
sinusoidal motion of the fish body, was principally used to
This work was supported by the STARFISH project of Defence Science
and Technology Agency (DSTA), Singapore, under Grant R-263-000-622-
232
Jian-Xin Xu is with Department of Electrical and Computer Engineer-
ing, Faculty of Engineering, National University of Singapore, Singapore
117576. He is also with NUS Graduate School for Integrative Sciences and
Engineering (NGS), National University of Singapore, Singapore 117456.
elexujx@nus.edu.sg
Xue-Lei Niu is with Department of Electrical and Computer Engineer-
ing, Faculty of Engineering, National University of Singapore, Singapore
117576. niuxuelei@nus.edu.sg
Guo Zhaoqin is with NUS Graduate School for Integrative Sciences and
Engineering (NGS), National University of Singapore, Singapore 117456,
guozhaoqin@nus.edu.sg
978-1-4577-2067-3/12/$26.00 c 2011 IEEE 308
study steady state propulsion. [9] designed a four-link robotic
fish, and implemented a PID controller and a fuzzy logic
controller to control its speed and orientation respectively.
[10] presented a neuronal model and a mechanical model of
fish swimming, and combined the two models together by
the transformation of the motoneuron activity to mechanical
forces and feedback of fish movements to stretch receptors.
Based on quasi-steady fluid flow theory, fish’s propulsion
model was established in [11], [12], and nonlinear control
method was investigated for trajectory stabilization of the
robotic fish.
Due to the complexity of modeling interaction force
between fish body and water, most existing works considered
steady speed of the robotic fish [7], [8], remaining dynamic
process not handled. Moreover, parameter uncertainties in
the fish model are paid little attention in existing works. In
this paper, we present a links-and-joints based robotic fish
model. Motion dynamics is given from previous result, and
two kinds of control laws are given. The main contributions
are: 1) By designing paticular reference joint angles, forward
locomotion of Carangiform fish is achieved; 2) Parameter
uncertainties and external disturbances are handled by sliding
mode control.
The rest of the paper is organized as follows. In Section
II, the mechanical model of the robotic fish and dynamics
of the system are given. Section III gives formulation of the
actual system which contains uncertainties and disturbances;
moreover, sliding mode controller design is given; In Sec-
tion IV, numerical examples are carried out to validate the
effectiveness of the controller. Section V concludes the paper
with a brief summary.
II. DYNAMIC M ODEL OF THE ROBOTIC F ISH
In this section, an Carangiform fish (carps, mackerels)
model is given to study its motion and control problems.
From a biological perspective, in this swimming mode, only
the latter part of the fish body participates in large amplitude
undulations, while the amplitude of the undulation of the
former part is small, which is different from the way that
Anguilliform fish moves. The most remarkable characteristic
in moving process of Carangiform fish, is that there exists
a body wave, traveling from head to tail [7]. Obviously, the
traveling direction of the body wave in the fish is backwards,
which is opposite to the direction that fish moves forward.
A. Fish Body Prototype
As shown in Fig. 1, we select the central line, which
locates at the center of the fish body and stretches from head
to tail, to represent the Carangiform fish, and we use links
 12th IEEE Workshop on Variable Structure Systems,
VSS’12, January 12-14, Mumbai, 2012
and joints to mimic its shape. The fish consists of 𝑁 links
and 𝑁 − 1 joints, where two connective links are connected
by one joint. There is one motor on each joint, and it exerts
torques to its neighbor links.
Fig. 1 shows the top view of the central curve of the
Carangiform fish. 𝑥𝑜𝑦 is the world coordinates system. The
position and orientation of each link 𝑖 are described by three
coordinates 𝑥𝑖 , 𝑦𝑖 and 𝜙𝑖 : 𝑥𝑖 and 𝑦𝑖 denote the position of the
midpoint of link 𝑖, while 𝜙𝑖 denotes the angle from +𝑥-axis
to link 𝑖. The links are numbered from head to tail (see Fig.
1b). Each link 𝑖 is impacted by two types of external forces:
hydrodynamic forces 𝑤𝑖 and torques 𝜏𝑖 , 𝜏𝑖−1 (see Fig. 2).
(a) The position(𝑥𝑖 , 𝑦𝑖 ) and orientation 𝜙𝑖 of each link 𝑖
(b) Numbering of links
Fig. 1. Sketch map of the Carangiform robotic fish model. (a) Position
and orientation representation. (b) Link numbering.
Fig. 2. External forces acting on link 𝑖.
B. Hydrodynamic Force
When there is relative motion between the fish and the
surrounding fluid, fluid is displaced and hydrodynamic force
arise. The force can be obtained through surface integrals
of vector force per area around the fish body. Since this
force is related with the geometry of the object immersed
978-1-4577-2067-3/12/$26.00 c 2011 IEEE
in water and relative velocity between the object and water,
in principle, the exact force distribution can be obtained by
solving the Navier-Stokes equation. However, the calculation
is quite complicated and time consuming [6]. In [10], the
force is simplified and easy to compute. As shown in Fig.
2, we adopt a simplified approximation of this force as (1)
and (2) indicate
𝑤𝑖⊥ = −𝑓𝑖⊥ (𝑣𝑖⊥ )2 𝑠𝑔𝑛(𝑣𝑖⊥ ) (1)
𝑤𝑖∥ = −𝑓𝑖∥ (𝑣𝑖∥ )2 𝑠𝑔𝑛(𝑣𝑖∥ ) (2)
where 𝑣𝑖⊥ , 𝑣𝑖∥ are parallel component and perpendicular
component of the velocity 𝑣𝑖 , and 𝑓𝑖⊥ , 𝑓𝑖∥ are the water
resistance coefficients in corresponding directions. The nota-
tion 𝑠𝑔𝑛(⋅) represents +1 if the element in the parentheses is
positive or −1 if negative. Based on the geometric relation-
ship (refer to Fig. 2), we have 𝑣𝑖⊥ = −𝑣𝑖𝑥 sin 𝜙𝑖 +𝑣𝑖𝑦 cos 𝜙𝑖 ,
𝑣𝑖∥ = 𝑣𝑖𝑥 cos 𝜙𝑖 +𝑣𝑖𝑦 sin 𝜙𝑖 , 𝑤𝑖𝑥 = −𝑤𝑖⊥ sin 𝜙𝑖 +𝑤𝑖∥ cos 𝜙𝑖 ,
𝑤𝑖𝑦 = 𝑤𝑖⊥ cos 𝜙𝑖 + 𝑤𝑖∥ sin 𝜙𝑖 , where 𝑣𝑖𝑥 , 𝑣𝑖𝑦 are projection
of the velocity 𝑣𝑖 on 𝑥-axis and 𝑦-axis; 𝑤𝑖𝑥 , 𝑤𝑖𝑦 are projec-
tion of the hydrodynamic force 𝑤𝑖 on 𝑥-axis and 𝑦-axis. All
of them are scalars. Hydrodynamic forces experienced by all
the links can be calculated the same way.
Since the link velocity 𝑣𝑖 can be possibly in any direction,
it is arduous to find each water resistance coefficient 𝑓 in
corresponding direction. Fortunately, 𝑓 remains unchanged
in the direction of parallelling the link, as well as in the per-
pendicular direction. Thus, we calculate the hydrodynamic
forces in such a way that the need of the value of 𝑓 in
arbitrary direction is avoided.
C. Lagrangian Formulation of the Mechanical Model
In this part, we give dynamics of the fish model. Details
of derivation see [1].
First, we define coordinates vector p ∈ ℜ3𝑁 as
p = [𝑥1 , 𝑦1 , 𝜙1 , 𝑥2 , 𝑦2 , 𝜙2 , ⋅ ⋅ ⋅ , 𝑥𝑁 , 𝑦𝑁 , 𝜙𝑁 ]𝑇
where the notation (⋅)𝑇 denotes transpose of a vector or a
matrix (⋅). The constraints in the system can be formulated
in matrix form
𝑥1 + 𝑙21 cos 𝜙1 − 𝑥2 + 𝑙22 cos 𝜙2
⎡ ⎤
⎢
⎢ 𝑦1 + 𝑙21 sin 𝜙1 − 𝑦2 + 𝑙22 sin 𝜙2 ⎥
⎥
⎢
⎢ 𝑥2 + 2 cos 𝜙2 − 𝑥3 + 2 cos 𝜙3
𝑙2 𝑙3 ⎥
⎥
g(p) = ⎢
⎢
⎢ 𝑦 2 + 𝑙2
2 sin 𝜙 2 − 𝑦 3 + 𝑙3
2 sin 𝜙 3 ⎥= 0
⎥
.. ⎥
.
⎢ ⎥
⎢ ⎥
⎣ 𝑥𝑁 −1 + 𝑙𝑁2−1 cos 𝜙𝑁 −1 − 𝑥𝑁 + 𝑙2𝑁 cos 𝜙𝑁 ⎦
⎢ ⎥
𝑦𝑁 −1 + 𝑙𝑁2−1 sin 𝜙𝑁 −1 − 𝑦𝑁 + 𝑙2𝑁 sin 𝜙𝑁
where 𝑙𝑖 is the length of link 𝑖, g(p) ∈ ℜ2(𝑁 −1) .
Define 𝐽(p) as the Jacobian of the constraints matrix g(p)
∂g(p)
𝐽(𝑝) =
∂p
where 𝐽(p) ∈ ℜ2(𝑁 −1)×3𝑁 . Afterwards, we give the external
forces vector which acts on individual coordinate of p
w = [𝑤1𝑥 , 𝑤1𝑦 , 𝜏1 , 𝑤2𝑥 , 𝑤2𝑦 , 𝜏2 − 𝜏1 , ⋅ ⋅ ⋅ ,
𝑤𝑁 𝑥 , 𝑤𝑁 𝑦 , −𝜏𝑁 −1 ]𝑇 (3)
309
 12th IEEE Workshop on Variable Structure Systems,
VSS’12, January 12-14, Mumbai, 2012
where 𝜏𝑖 − 𝜏𝑖−1 represents the total torque exerted on link
𝑖. It is worth noting that 𝜏0 = 𝜏𝑁 = 0, since there are no
torques at the endpoints.
The system dynamics is then given as
p̈ = 𝐴(p)ṗ + 𝐵(p)w
where 𝐴(p) = −𝑀 −1 𝐽 𝑇 (𝐽𝑀 −1 𝐽 𝑇 )−1 𝐽, ˙ 𝐵(p) =
𝑀 −1 [𝐼 − 𝐽 𝑇 (𝐽𝑀 −1 𝐽 𝑇 )−1 𝐽𝑀 −1 ], 𝐴(p) ∈ ℜ3𝑁 ×3𝑁 ,
𝐵(p) ∈ ℜ3𝑁 ×3𝑁 , 𝑀 ∈ ℜ3𝑁 ×3𝑁 is the mass matrix and
it can be written as
𝑀 = 𝑑𝑖𝑎𝑔{𝑚1 , 𝑚1 , 𝐼1 , 𝑚2 , 𝑚2 , 𝐼2 , ⋅ ⋅ ⋅ , 𝑚𝑁 , 𝑚𝑁 , 𝐼𝑁 }
where 𝑚𝑖 is the mass and 𝐼𝑖 is the moment of inertia of
link 𝑖. The notation 𝑑𝑖𝑎𝑔{} represents that 𝑀 is a diagonal
matrix, and the diagonal elements are in the braces. 𝐼 is
identity matrix with the same dimension as 𝑀 .
(4) contains all the acceleration terms, of which we
are more interested in angular acceleration terms 𝜙¨𝑖 . By
partitioning (4), we obtain dynamics of 𝜙
𝜙¨ = 𝐴1 (p)ṗ + 𝐵1 (p)w𝑥 + 𝐵2 (p)w𝑦 + 𝐵3 (p)𝐵𝜏 𝜏
= [𝜙1 , 𝜙2 , ⋅ ⋅ ⋅ , 𝜙𝑁 ]𝑇
where 𝜙 v⊥sin = [𝑣1⊥ 𝑠𝑔𝑛(𝑣1⊥ ) sin 𝜙1 , ⋅ ⋅ ⋅ , 𝑣𝑁
w𝑥 = [𝑤1𝑥 , 𝑤2𝑥 , ⋅ ⋅ ⋅ , 𝑤𝑁 𝑥 ]𝑇
w𝑦 = [𝑤1𝑦 , 𝑤2𝑦 , ⋅ ⋅ ⋅ , 𝑤𝑁 𝑦 ]𝑇
= [𝜏1 , 𝜏2 , ⋅ ⋅ ⋅ , 𝜏𝑁 −1 ]𝑇
𝜏 𝑇
1 0 ⋅⋅⋅ 0
⎡ ⎤
. ..
⎢ −1 1
⎢ .. .
⎥
⎥
⎢ ⎥
𝐵𝜏 = ⎢ 0 −1 . . . 0
⎢ ⎥
⎥
⎢ ⎥
⎢ . . .
⎣ .. .. ..
⎥
1 ⎦
0 ⋅⋅⋅ 0 −1
and 𝐴1 (p)∈ℜ𝑁 ×3𝑁 , 𝐵1 (p)∈ℜ𝑁 ×𝑁 , 𝐵2 (p)∈ℜ𝑁 ×𝑁 , 𝐵3 (p)∈
ℜ𝑁 ×𝑁 are corresponding coefficient matrices obtained from
matrix 𝐴(p), 𝐵(p) in (4). It is worth noting that the
dimension of 𝜏 is 𝑁 − 1, one less than the total number
of links 𝑁 .
III. S LIDING M ODE C ONTROL D ESIGN FOR THE
ROBOTIC F ISH S YSTEM
Modeling inaccuracies always exist and have strong ad-
verse effects on control systems. Thus, any practical design
must address them explicitly [13]. Otherwise, the control law
may lose effect since the actual parameters deteriorate the
performance of the whole system. Here we adopt sliding
mode control, which belongs to robust controllers. In this
robotic fish system, it is obvious that the number of actuators
is less than that of reference input. Therefore, we have some
considerations on that when we design the sliding mode
control law.
978-1-4577-2067-3/12/$26.00 c 2011 IEEE
A. Parameter Uncertainties
In the robotic fish model we construct, many parameters
involve uncertainties. These uncertainties either come from
inaccuracy in the modeling, or come from unpredictable
influence of surroundings. The water resistance coefficient 𝑓
(4) can be affected by many factors, such as different velocities
of fish with respect to the environment. Thus, it is rather an
estimated parameter than an accurate one. The mass matrix
𝑀 can be measured accurately on ground, but when the fish
comes into water, 𝑀 becomes inaccurate because of added
mass effect. Though we cannot know the exact information
of the interested parameters due to complex factors, those
parameters always change in a predictable range. This is
reasonable because every parameter has its own physical
meaning, thus it neither blows up to infinity nor becomes
too small. Then, we can always give an upper bound and a
lower bound for each parameter.
We define that
𝐹⊥ = 𝑑𝑖𝑎𝑔{𝑓1⊥ , 𝑓2⊥ , ⋅ ⋅ ⋅ , 𝑓𝑁 ⊥ }
𝐹∥ = 𝑑𝑖𝑎𝑔{𝑓1∥ , 𝑓2∥ , ⋅ ⋅ ⋅ , 𝑓𝑁 ∥ }
Assume that there exist parameter uncertainties on 𝑀 , 𝐹⊥
(5)
and 𝐹∥ , and their norms are bounded.
For derivation convenience, we define that
2 2
⊥ 𝑠𝑔𝑛(𝑣𝑁 ⊥ ) sin 𝜙𝑁 ]
𝑇
2 2
v⊥cos = [𝑣1⊥ 𝑠𝑔𝑛(𝑣1⊥ ) cos 𝜙1 , ⋅ ⋅ ⋅ , 𝑣𝑁 ⊥ 𝑠𝑔𝑛(𝑣𝑁 ⊥ ) cos 𝜙𝑁 ]
𝑇
2 2
v∥sin = [𝑣1∥ 𝑠𝑔𝑛(𝑣1∥ ) sin 𝜙1 , ⋅ ⋅ ⋅ , 𝑣𝑁 ∥ 𝑠𝑔𝑛(𝑣𝑁 ∥ ) sin 𝜙𝑁 ]
𝑇
2
v∥cos = [𝑣1∥ 2
𝑠𝑔𝑛(𝑣1∥ ) cos 𝜙1 , ⋅ ⋅ ⋅ , 𝑣𝑁 ∥ 𝑠𝑔𝑛(𝑣𝑁 ∥ ) cos 𝜙𝑁 ]
Then, hydrodynamic forces w𝑥 and w𝑦 can be written as
w𝑥 = 𝐹⊥ v⊥sin − 𝐹∥ v∥cos (6)
w𝑦 = −𝐹⊥ v⊥cos − 𝐹∥ v∥sin
(7)
where w𝑥0 , w𝑦0 are nominal values of w𝑥 , w𝑦 . From a
practical point of view, the coordinate p and its derivative
ṗ are always bounded, which indicates that both 𝐴(p) and
𝐵(p) are bound, because 𝐴(p) and 𝐵(p) are functions of p
and ṗ. Together with the fact that 𝐹⊥ and 𝐹∥ are bounded,
we can know that w𝑥 and w𝑦 are bounded from (6) and (7).
B. External Disturbances
Besides parameter uncertainties, we have to take into ac-
count environmental disturbances when considering system
dynamics. Disturbances are always existing in the robotic
fish system, no matter in outdoor or indoor environments.
In outdoor environment, such as seas or rivers, the robotic
fish will experience unpredictable currents or waves, which
are generated by winds or heat exchange, either beneath or
on the surface of the water. In indoor environment, such as
a water tank, the robotic fish will experience the reflection
wave from the water container, which are generated by the
swimming motion of the fish itself.
In general, these disturbances can be both additive and
multiplicative to the dynamic equations of motion. However,
the additive disturbances are a good approximation for most
310
 12th IEEE Workshop on Variable Structure Systems,
VSS’12, January 12-14, Mumbai, 2012
marine control applications [14]. Hence, we assume that the
disturbances have additive format in this work.
The disturbances function as external forces, which are
similar to w𝑥 , w𝑦 and 𝜏 in (5). Since the disturbances adhere
to the original external forces and they are additive, (5) can
be written as follows
𝜙¨ = 𝐴1 (p)ṗ + 𝐵1 (p)(w𝑥 + d𝑥 ) + 𝐵2 (p)(w𝑦 + d𝑦 )
+𝐵3 (p)𝐵𝜏 (𝜏 + d𝜏 )
where d𝑥 ∈ ℜ𝑁 and d𝑦 ∈ ℜ𝑁 are the disturbance vectors in
the directions of 𝑥-axis and 𝑦-axis, respectively, d𝜏 ∈ ℜ𝑁 −1
is the disturbance vector on the torque 𝜏 . In practice, d𝑥 , d𝑦
and d𝜏 will change in a reasonable range.
Note that among the parameter uncertainties and external
disturbances that appear in (8), only d𝜏 is matched, while all
the others are unmatched, because they are not in the range
space of the input matrix 𝐵3 (p)𝐵𝜏 .
C. Sliding Mode Control Law Design
Generally, there are two standard steps in sliding mode
control design: 1) a sliding surface is given such that system
on it manifests desired behavior; 2) a discontinuous control
law is utilized to drive the system states into that surface and
stay on it for all future time [15].
The sliding mode control law is composed of two parts.
The first part is used to handle the nominal model, while the
second is used to handle system uncertainties. Since there
are not enough number of actuators to track all the reference
inputs, we have to make a trade-off when setting the control
objective, i.e., tracking the same number of reference inputs
as that of actuators. In this model, the number of actuators,
i.e., the number of torques, applied on the fish is 𝑁 −1. Thus,
we make the dimension of the sliding surface be 𝑁 − 1.
Before designing the sliding surface, we first define angu-
lar error and its derivative
e = 𝜙𝑁 𝑒𝑤 − 𝜙𝑟𝑁 𝑒𝑤 ė = 𝜙˙ 𝑁 𝑒𝑤 − 𝜙˙ 𝑟𝑁 𝑒𝑤
where 𝜙𝑁 𝑒𝑤 = [𝜙1 , 𝜙2 , ⋅ ⋅ ⋅ , 𝜙𝑁 −1 ]𝑇 , representing the
first 𝑁 − 1 actual joint angles, and 𝜙𝑟𝑁 𝑒𝑤
[𝜙1𝑟 , 𝜙2𝑟 , ⋅ ⋅ ⋅ , 𝜙(𝑁 −1)𝑟 ]𝑇 , representing the first 𝑁 − 1 refer-
ence joint angles. The dynamics of 𝜙𝑁 𝑒𝑤 is
𝜙¨𝑁 𝑒𝑤 = 𝐴2 (p)ṗ + 𝐵4 (p)(w𝑥 + d𝑥 ) + 𝐵5 (p)(p)(w𝑦 + d𝑦 )
+𝐵6 (p)𝐵𝜏 (𝜏 + d𝜏 )
= 𝐴2𝑛 (p)ṗ + 𝐵4𝑛 (p)w𝑥𝑛 + 𝐵5𝑛 (p)w𝑦𝑛
+𝐵6𝑛 (p)𝐵𝜏 𝜏 + d(p, 𝜏, 𝑡)
where 𝐴2 (p), 𝐵4 (p), 𝐵5 (p), 𝐵6 (p), are submatrices ob-
tained from matrix 𝐴(p), 𝐵(p) in (4), corresponding to
𝜙¨𝑁 𝑒𝑤 , and 𝐴2𝑛 (p), 𝐵4𝑛 (p), 𝐵5𝑛 (p), 𝐵6𝑛 (p) are their nom-
inal values. d(p, 𝜏, 𝑡) = (𝐴2 − 𝐴2𝑛 )ṗ + (𝐵4 w𝑥 − 𝐵4𝑛 w𝑥𝑛 +
𝐵4 d𝑥 ) + (𝐵5 w𝑦 − 𝐵5𝑛 w𝑦𝑛 + 𝐵5 d𝑦 ) + (𝐵6 𝐵𝜏 𝜏 − 𝐵6𝑛 𝐵𝜏 𝜏 +
𝐵6 𝐵𝜏 d𝜏 ), represents the difference between the actual terms
and nominal terms. In the expression of d(p, 𝜏, 𝑡), all the
terms are bounded, thus we assume that the norm of it has
an upper bound
∣∣d(p, 𝜏, 𝑡)∣∣ ≤ 𝑑max
978-1-4577-2067-3/12/$26.00 c 2011 IEEE
From the definition of e, we set the control objective as
tracking the first 𝑁 − 1 reference inputs. Next, we define the
sliding surface as
𝜎 = 𝐶e + ė (10)
where 𝐶 is a diagonal matrix whose entries are positive
scalars.
(8) Assume that information of the coordinate vector p and
its velocity ṗ is available by means of vision or other
measurement system. Now, we give the control law. As stated
before, the control law consists of two parts
𝜏 = 𝜏0 + 𝜏𝑠 (11)
−1 ¨ ˙ ˙
𝜏0 = (𝐵6𝑛 𝐵𝜏 ) [𝜙𝑟𝑁 𝑒𝑤 − 𝐶(𝜙𝑁 𝑒𝑤 − 𝜙𝑟𝑁 𝑒𝑤 )
−(𝐴2𝑛 ṗ + 𝐵4𝑛 w𝑥𝑛 + 𝐵5𝑛 w𝑦𝑛 )] (12)
𝜎
𝜏𝑠 = −𝜌(𝐵6𝑛 𝐵𝜏 )−1 (13)
∣∣𝜎∣∣
where 𝜌 = 𝑑max + 𝜂, 𝜂 is a positive constant. 𝜏0 is used to
handle nominal model, 𝜏𝑠 is used to handle the uncertainties.
Then, we have the following theorem.
Theorem 1: Consider the nonlinear system (9) associated
with the chosen sliding surface 𝜎 = 0. Under the control law
(11)-(13), the sliding surface will be reached in finite time.
Proof. First we define the Lyapunov function
1 𝑇
𝑉 = 𝜎 𝜎
2
Differentiating it, we obtain
𝑉˙ = 𝜎 𝑇 𝜎˙ = 𝜎 𝑇 (𝐶 ė + ë)
= 𝜎 𝑇 [𝐶(𝜙˙ 𝑁 𝑒𝑤 − 𝜙˙ 𝑟𝑁 𝑒𝑤 ) − 𝜙¨𝑟𝑁 𝑒𝑤 + (𝐴2 ṗ + 𝐵4 w𝑥
+𝐵4 d𝑥 + 𝐵5 w𝑦 + 𝐵5 d𝑦 + 𝐵6 𝐵𝜏 𝜏 + 𝐵6 𝐵𝜏 d𝜏 )]
= 𝜎 𝑇 [𝐶(𝜙˙ 𝑁 𝑒𝑤 − 𝜙˙ 𝑟𝑁 𝑒𝑤 ) − 𝜙¨𝑟𝑁 𝑒𝑤
+(𝐴2𝑛 ṗ + 𝐵4𝑛 w𝑥𝑛 + 𝐵5𝑛 w𝑦𝑛 + 𝐵6𝑛 𝐵𝜏 𝜏 ) + d]
Substituting (11), (12) and (13) into 𝑉˙ , one obtains
=
𝑉˙ = −𝜌∣∣𝜎∣∣ + 𝜎 𝑇 d ≤ −𝜌∣∣𝜎∣∣ + ∣∣𝜎∣∣ ⋅ ∣∣d∣∣
≤ −𝜌∣∣𝜎∣∣ + 𝑑max ∣∣𝜎∣∣ = −𝜂∣∣𝜎∣∣
It is obvious that 𝑉˙ is negative definite. By Lyapunov
theorem for stability, the equilibrium at the origin 𝜎 = 0 is
asymptotically stable. If 𝜙𝑁 𝑒𝑤 (𝑡 = 0) is off 𝜙𝑟𝑁 𝑒𝑤 (𝑡 = 0)
in the beginning, the sliding surface is reached in a finite
(9) time 𝑡𝑟𝑒𝑎𝑐ℎ ≤ ∥𝜎(𝑡 = 0)∥∞ /𝜂, where ∥(⋅)∥∞ denotes the
∞−norm of (⋅). After the system reaches the sliding surface
𝜎 = 0, it stays there. In the sliding mode, 𝜎(𝑡) = 0, 𝜎(𝑡)
˙ =
0, the equivalent control is 𝜏𝑒𝑞 = (𝐵6 𝐵𝜏 )−1 [𝜙¨𝑟𝑁 𝑒𝑤 −
𝐶(𝜙˙ 𝑁 𝑒𝑤 − 𝜙˙ 𝑟𝑁 𝑒𝑤 )−(𝐴2 ṗ+𝐵4 w𝑥 +𝐵4 d𝑥 +𝐵5 w𝑦 +𝐵5 d𝑦 +
𝐵6 𝐵𝜏 d𝜏 )]. (10) gives the dynamics of e, which contains the
first three angular errors. Since all entries of the diagonal
matrix 𝐶 are chosen to be positive scalars, it is easy to show
that on the sliding surface, each single element of e always
converges to 0, thus yielding the result that the first 𝑁 − 1
reference inputs can be tracked.
311
 12th IEEE Workshop on Variable Structure Systems,
VSS’12, January 12-14, Mumbai, 2012

IV. N UMERICAL E XAMPLES d𝜏 = 10−3 × [0.2 + 0.2 sin 𝑡, 0.1 − 0.2 cos 𝑡]𝑇 . Chattering
Forward motion is the most common locomotion pattern phenomenon always exists in sliding mode control, which
of Carangiform fish (in this work, all the examples are given is a character of it. To have a smoother control signal, we
𝜎
by using forward locomotion). If the fish moves forward, replace ∥𝜎∥ in (13) with a saturation function 𝑠𝑎𝑡(𝜎).
there exists a body wave traveling backwards. Since the wave ⎧ 𝜎
travels from its head to tail, the head is preceding the tail
 , if ∣∣𝜎∣∣ > 𝜖,
∣∣𝜎∣∣
⎨
affected by the wave. More generally, the movement of the 𝑠𝑎𝑡(𝜎) =
⎩ 𝜎,
 otherwise.
former part of the body has a phase lead than the latter one, 𝜖
and it is reflected in the phase difference among the link
and here we choose 𝜖 = 0.1
orientation angle 𝜙𝑗 (𝑗 = 1, ⋅ ⋅ ⋅ , 𝑁 ). For the reason that the
In the first scenario, we use sliding mode control, and se-
backward moving wave has the same oscillating frequency
lect the parameters as 𝐶 = 𝑑𝑖𝑎𝑔{10, 1} (a diagonal matrix),
at different places, we suppose each 𝜙𝑗 follow the same
𝑑max = 40, 𝜂 = 0.1. At time 𝑡 = 0, the fish is still, and its
angular frequency. Following these considerations, we let the
three links are aligned on 𝑥-axis with its head at the origin,
reference 𝜙𝑗𝑟 assume the following form
which means p = [0.15 0 0 0.45 0 0 0.75 0 0 1.05 0 0]𝑇 .
𝜙𝑗𝑟 = 𝐴𝑚 (𝑗) ⋅ sin[𝜔𝑡 + (2 − 𝑗)𝜃] (14) The control torques are shown in Fig. 3.
where 𝑗 = 1, 2, ⋅ ⋅ ⋅ , 𝑁 . 𝑡 denotes time instant. 𝐴𝑚 (𝑗), 𝜔 are
the amplitude and angular frequency of 𝜙𝑗𝑟 respectively, and 0.06
τ1 τ2

𝜃 is the phase lead of link 𝑖 comparing with its latter one. 0.04

Note that in Carangiform fish, the amplitude of the body

0.02
wave gradually increases from its head to its tail, thus we
have the following relation: 𝐴𝑚 (1) < ⋅ ⋅ ⋅ < 𝐴𝑚 (𝑗) < ⋅ ⋅ ⋅ < 0

𝐴𝑚 (𝑁 ). Torques
In this model, we select 𝑁 = 3, i.e., the robotic fish
−0.02

consists of three links. The first link is the longest, and the −0.04

third link is the shortest. Their length proportion is referred to −0.06

[2]. Table I shows mechanical parameters of the links, where

𝑙𝑖 , 𝑚𝑖 , 𝐼𝑖 are the length, mass and moment of inertia of link −0.08

𝑖 respectively, 𝑓𝑖⊥ and 𝑓𝑖∥ are estimated water resistance −0.1

coefficients [10]. Their SI units are 𝑚(𝑚𝑒𝑡𝑒𝑟), 𝑘𝑔, 𝑘𝑔 ⋅ 𝑚2 ,

0 2 4 6 8 10 12 14 16 18 20
Time (s)

𝑁 𝑠2 /𝑚2 , 𝑁 𝑠2 /𝑚2 .
Fig. 3. Scenario 1: Torques trajectory (Sliding mode control).
TABLE I
M ECHANICAL PARAMETERS OF THE L INKS
Link # 𝑙𝑖 𝑚𝑖 𝐼𝑖 𝑓𝑖⊥ 𝑓𝑖∥
1. 0.15 0.45 1.4603 × 10−3 2.7 2.70 0.1

2. 0.108 0.168 2.721 × 10−4 5.4 1.40 0

3. 0.0975 0.1236 1.632 × 10−4 5.4 1.141
−0.1

−0.2

Based on (14), we give the reference angles 𝜙𝑗𝑟 . Since this −0.3
x1 trajectory

robotic fish is essentially an underactuated system, tracking

−0.4
of arbitrary number of reference inputs is impossible. How-
−0.5
ever, by reference planning approach in [1], which conducts
equilibrium analysis at the neighborhood of the equilibrium −0.6

point, this problem can be handled. Following this method −0.7

and the relation between the amplitudes of the body wave, −0.8

we choose 𝜃 = 2.003, 𝐴𝑚 (1) = 0.3142, 𝐴𝑚 (2) = 0.3227, −0.9

0 2 4 6 8 10 12 14 16 18 20
𝐴𝑚 (3) = 1.0090. Other parameter is appropriately chosen Time (s)

as 𝜔 = 2𝜋. In this section, the above set of parameters

(𝜃, 𝐴𝑚 , 𝜔), applies to all scenarios, and we only consider Fig. 4. Scenario 1: 𝑥1 trajectory (Sliding mode control).
the forward moving case due to space limitation.
For simplicity, we suppose that the parameter uncer- In the second scenario, we use computed torque control
tainties are in the following form: 𝑀 = (1 + 𝛼)𝑀0 , method [1] in (16), with the feedback coefficients 𝑘1 = 10,
𝐹⊥ = (1 + 𝛽1 )𝐹⊥0 , 𝐹∥ = (1 + 𝛽2 )𝐹∥0 , where 𝛼 = 𝑘2 = 1.
0.2, 𝛽1 = 0.2, 𝛽2 = 0.2. The disturbances parameters
are: d𝑥 = 10−3 × [1 + sin 𝑡, −2 + cos 𝑡, 0.5 − 2 sin 𝑡]𝑇 , 𝜏 = (𝐵𝜏𝑇 𝐵3 𝐵𝜏 )−1 𝐵𝜏𝑇 [𝜙¨𝑟 + 𝑘1 (𝜙𝑟 − 𝜙) + 𝑘2 (𝜙˙ 𝑟 − 𝜙)
˙
d𝑦 = 10−3 × [0.2 + 0.3 sin 𝑡, 1 + 2 cos 𝑡, 0.3 − 4 sin 𝑡]𝑇 , −(𝐴1 ṗ + 𝐵1 w𝑥 + 𝐵2 w𝑦 )] (16)

978-1-4577-2067-3/12/$26.00 c 2011 IEEE 312

 12th IEEE Workshop on Variable Structure Systems,
VSS’12, January 12-14, Mumbai, 2012

where 𝜙𝑟 = [𝜙1𝑟 , 𝜙2𝑟 , ⋅ ⋅ ⋅ , 𝜙𝑁 𝑟 ]𝑇 is the vector of reference
angles for all the links, and 𝑘1 , 𝑘2 are coefficients relating to
feedback terms. Here we choose 𝑘1 = 10, 𝑘2 = 1. Other
parameters and initial condition are the same as the first
scenario. The comparison of angular error between sliding
mode controller and computed torque controller is shown
in Fig. 5. It is obvious that by SMC method, the first two
joint angular errors quickly converge to 0 after a short period
of time. The third joint angular error changes around 0 but
cannot converge. While by computed torque method, the
three joint angular errors are much larger than those obtained
from SMC method, and none of them converge to 0 in the
end.
0.4
SMC
trajectory
which means arbitrary number of trajectory tracking is
impossible. Furthermore, robustness can not be guaranteed
for unmatched uncertainties or disturbances.
V. C ONCLUSION
In this paper, the dynamic model of the Carangiform
robotic fish is first given. Based on the fact that body wave
exists on the body of a traveling fish, we give the reference
angles for all the links’ orientation by using sinusoidal
functions. We find that when the former reference angle has
a phase lead compared with the latter one, the fish moves
forward. Sliding mode control is proposed to handle the
actual system model, which takes into account parameter
uncertainties and external disturbances. SMC design can
promise theoretically that angular error of the first two links
converge to zero. Numerical results show that the effec-
tiveness of SMC to resist uncertainties, and better tracking

CTC
0.2 performance is obtained compared with that using computed
0
torque control.
1r

R EFERENCES
φ −φ

−0.2
1

−0.4
[1] J.-X. Xu and X.-L. Niu, “Analytical control design for a biomimetic
0 2 4 6 8 10 12 14 16 18 20 robotic fish,” in Industrial Electronics (ISIE), 2011 IEEE International
Time (s) Symposium on, June 2011, pp. 864 –869.
0.4
[2] J. Yu, L. Wang, and M. Tan, “Geometric optimization of relative link
SMC lengths for biomimetic robotic fish,” IEEE Transactions on Robotics,
trajectory

0.2 CTC vol. 23, no. 2, pp. 382–386, Apr. 2007.

0
[3] C. Zhou, M. Tan, Z. Cao, S. Wang, D. Creighton, N. Gu, and
S. Nahavandi, “Kinematic modeling of a bio-inspired robotic fish,”
−0.2 in IEEE International Conference on Robotics and Automation, 2008.
2r
φ −φ

−0.4
ICRA 2008., May 2008, pp. 695 –699.
2

[4] K. Zou, C. Wang, G. Xie, T. Chu, L. Wang, and Y. Jia, “Cooperative

−0.6 control for trajectory tracking of robotic fish,” in American Control
0 2 4 6 8 10 12 14 16 18 20
Time (s) Conference, 2009. ACC ’09., June 2009, pp. 5504 –5509.
[5] M. Sfakiotakis, D. Lane, and J. Davies, “Review of fish swimming
1.5 modes for aquatic locomotion,” IEEE Journal of Oceanic Engineering,
SMC
trajectory

1 CTC vol. 24, no. 2, pp. 237 –252, Apr. 1999.

[6] J. Colgate and K. Lynch, “Mechanics and control of swimming: A
0.5 review,” IEEE Journal of Oceanic Engineering, vol. 29, no. 3, pp.
0 660 – 673, July 2004.
3r

[7] M. J. Lighthill, “Aquatic animal propulsion of high hydromechanical

φ −φ

−0.5 efficiency,” Journal of Fluid Mechanics, vol. 44, no. Nov., pp. 265–
3

−1
0 2 4 6 8 10 12 14 16
Time (s)
Fig. 5. Comparison of angular error between Scenario 1: sliding mode
control (SMC) and Scenario 2: computed torque control (CTC).
Comparing the two scenarios, we find that when there exist
parameter and unmodeled uncertainties, computed torque
control cannot work very well, because the error between
the actual joint angles and reference joint angles are large
and always exists. When we use sliding mode control, the
first two reference inputs can be perfectly tracked, thus we
achieve the goal in Sec. IV-B when designing the sliding
surface.
In Scenario 1, though 𝜙1𝑟 and 𝜙2𝑟 can be tracked ac-
curately, perfect tracking of 𝜙3𝑟 cannot be promised the-
oretically. The fundamental reason is that the system is
underactuated, i.e., the number of actuators in the system
is fewer than the number of independent physical variables,
301, 1970.
18 20 [8] ——, “Large-amplitude elongated-body theory of fish locomotion,”
Proceedings of the Royal Society of London. Series B, Biological
Sciences, vol. 179, no. 1055, pp. 125–138, 1971.
[9] J. Z. Yu, M. Tan, S. Wang, and E. Chen, “Development of a biomimetic
robotic fish and its control algorithm,” IEEE Transactions on Systems
Man and Cybernetics Part B: Cybernetics, vol. 34, no. 4, pp. 1798–
1810, Aug. 2004.
[10] O. Ekeberg, “A combined neuronal and mechanical model of fish
swimming,” Biological Cybernetics, vol. 69, no. 5-6, pp. 363–374,
Oct. 1993.
[11] K. A. Morgansen, V. Duidam, R. J. Mason, J. W. Burdick, and R. M.
Murray, “Nonlinear control methods for planar carangiform robot
fish locomotion,” in IEEE International Conference on Robotics and
Automation, 2001. Proceedings 2001 ICRA., vol. 1, 2001, pp. 427 –
434 vol.1.
[12] K. A. Morgansen, P. A. Vela, and J. W. Burdick, “Trajectory stabiliza-
tion for a planar carangiform robot fish,” in Robotics and Automation,
2002. Proceedings. ICRA ’02. IEEE International Conference on,
vol. 1, 2002, pp. 756 – 762 vol.1.
[13] J. Slotine and W. Li, Applied nonlinear control. Prentice Hall, 1991.
[14] T. Fossen, Guidance and Control of Ocean Vehicles. New York, NY:
Wiley, 1995.
[15] W. Cao and J. Xu, “Nonlinear integral-type sliding surface for both
matched and unmatched uncertain systems,” IEEE Transaction on
Automatic Control, vol. 49, no. 8, pp. 1355–1360, Aug 2004.

978-1-4577-2067-3/12/$26.00 c 2011 IEEE 313