Multi-variable Constrained Control Approach for a Three-Dimensional

Eel-like Robot
M. El Rafei, M. Alamir, N. Marchand, M. Porez and F. Boyer

Abstract— In this paper, a multi-variable feedback design feedback that tracks some reference trajectory. Another in-
for the 3D movement of an eel-like robot is presented. Such a teresting approach was proposed in [10], [11], [16] where
robot is under construction in the context of a national French averaging formulas have been derived to describe the mean
robotic project. The proposed feedback enables the tracking of
a desired 3D position of the eel’s head as well as the stabilization behavior over an undulatory cycle. A design procedure for a
of the rolling angle. The control design is based on a recently biomimetic robot-fish based on improved kinematic propul-
developed reduced model that have been validated using a 3D sive model has been described in [17] where the basic motion
complete continuous model described in [3]. Several scenarios control laws were presented. For a detailed review of existing
are proposed to assess the efficiency of the proposed feedback works on the mechanics and the control of swimming, the
law.
reader can refer to [6]. However, few research have been
I. I NTRODUCTION

In this paper, current researches on the control of an

eel-like robot are presented. This work is carried out in
the context of a multidisciplinary French national research
project 1 . The aim of this project is to design, construct and
control the 3D motion of an eel-like robot. The prototype Fig. 1. The experimental assembly (under construction)
under construction is obtained by connecting many parallel
platforms (see figure 1). The eel’s body will then be covered
conducted on the control design for 3D motion of the eel-
by a deformable “skin” in order to achieve high performance
like robots. Results on the 3D control of an eel-like robot are
swimming. As it has been underlined by many researchers
presented in [1] and [7] where a complete control scheme
in this field, understanding the dynamics of such robots may
for 3D movement of the continuous model [3] was proposed.
be of a great interest in improving the manoeuvrability of
The motion and the velocity in the transverse plane are
under-water vehicles [15], [8], [11], [9], [6], [12], [14], [2].
controlled by monitoring the oscillatory gait characteristics
A 3D continuous model of the target prototype has been while the altitude changes and the rolling stabilization task
proposed in [3] using the geometrically exact theory of are handled by means of two pectoral fins that are attached to
beams under finite deformations [13]. This model that is not the eel’s head [1] or by 3D robot’s body movements without
suitable for use in advance control derivation was used to using pectorals fins [7]. These decoupled control laws suffer
identify and validate a low dimensional and computationally however from a high sensitivity to the choice of the control
efficient reduced mean model that can be used in advanced parameters. That is why a coupled multi-variable control are
control design. This 3D reduced model is clearly used here still to be developed.
to design and validate the proposed 3D control feedback.
In this work, a coupled control scheme for 3D movement
However, for the lack of space, the reduced model derivation
of the robot’s reduced model is proposed. Basically, given
is not treated by this paper. Nevertheless, it is still briefly
the robot state and the targeted position, a desired velocity is
described in section II.
computed that enables to incrementally achieve the tracking
There exist many works in the literature that studied the
mission. This desired velocity is then used to build a con-
eel-like robots movements. In [9] and the related works,
straint quadratic optimization problem in which, the decision
the 2D movement of an eel-like robot has been studied.
variable are incremental actions to be taken over the current
The rolling cart analogy is used in order to derive state
sampling period. This optimization problem is used also to
achieve roll angle stabilization.
This work was supported by the French National Center for Scientific
Research (CNRS) in the context of the ROBEA-project and the French This paper is organized as follows : First, the reduced
National research agency project (ANR-RAAMO). model is summarized in section II. Section III clearly states
M. El Rafei, M. Alamir and N. Marchand are the 3D control problem. The control strategy is then pre-
with the control system department of GIPSA-Lab
[email protected], sented in section IV, namely, the tracking of a desired 3D
[email protected], [email protected] position as well as the rolling angle stabilization. Some 3D
M. Porez and F. Boyer are with the Institut de Recherche scenarios are proposed in section V. The paper ends by some
en Cybernétique de Nantes. [email protected],
[email protected] concluding remarks together with the road map for future
1 http://www.irccyn.ec-nantes.fr/hebergement/ROBEA/ works.
   
II. T HE EEL - LIKE ROBOT REDUCED MODEL ẋ V1
 ẏ  = Rq (Q)  0  (10)
This section presents a simplified dynamic model for the
ż V3
3D Eel-like robot. This model is used for the control purpose
and it is based on the continuous model [3]. It consists in where :
modeling the eel’s head linear and angular mean velocities as • q1 is an internal variable.
dynamical functions of the control input. Figure 2 illustrates • ρ is the eel’s body curvature. Note that the control law
the basic notations used in the description of the model where applies a uniform additional curvature (uniform along
the following notations are used : the body but variable in time) that is added to the non
• (0, E1 , E2 , E3 ) denotes the earth frame assimilated to uniform curvature needed to enhance the undulation
a galilean reference. wave.
• (0, t10 , t20 , t30 ) refers to the mobile frame attached to • wr = ρV1 .


the eel’s head. • Q = q0 qx qy qz is the quaternion that
• t10 , t20 and t30 are respectively the head’s roll, pitch represents the head frame’s orientation with respect to
and yaw axes. the inertial frame. We can also represent this orientation
by a rotation matrix Rq (Q). The quaternion (and its


• V 0 = V1 V2 V3 denotes the mean linear velocity
of the eel’s time derivative) can be related to the rotation vector
R t head expressed in the head frame. Moreover
V0 = T1 t−T kV0 (τ )kdτ w. This relation (9) enables to have Q as a function
• T is the undulation period. Note that the 3D control of w. For more informations about quaternion related
of the eel-like robot is achieved here without use of formalism, the reader can refer to [5].
pectoral fins through the deformations of the end part  
of the deformable body [7]. The key idea consists in 0 −wp −wq −wr
 wp 0 wr −wq 
applying torsion and pitch movements to the tailing part • M (w) =   is a skew-
 wq −wr 0 wp 
of the body (see figure 3) in phase with its undulatory
movement in order to generate pressure wave and fluid wr wq −wp 0
symmetric tensor.
forces that are used by
the control law. • O(x, y, z) represents the 3D coordinates of the eel’s
• w0 = wp wq wr denotes the mean rotation vec-
head.
torR (angular velocity) of the head. Moreover w0 =
1 t • u = (up , uq , u1 , u2 , u3 ) is the control input (see section
T t−T kw0 (τ )kdτ III for more details).
• wp , wq and wr are respectively the head’s roll, pitch and ∞ ∞ ∞
• λ1 , V1 , λ3 , V3 , λp , p , λ1q , λ2q , µq are the identified
yaw mean velocities.
parameters as functions of the control vector u and the
leading velocity component V1 .
• λr is a fixed constant parameter that directly monitors
the body curvature.
This model is used here to derive the coupled 3D control for
the robot’s 3D motion.
III. S TATEMENT OF THE CONTROL PROBLEM
Recall that the reduced model is identified and validated
based on the complete continuous model presented in [3].
In the later, the robot is viewed as a beam defined by a
continuous assembly of rigid cross sections. For this model

Fig. 2. Frames and parametrization of the reduced eel-like robot model a vector field K(t, X) = K1 (t, X) K2 (t, X) K3 (t, X)
is the control input in which, t is the time, X designates
Our experience based on the continuous model [3] suggests
the material abscissa along the eel’s mean line. The last
to use the following structure for the reduced mean model :
two components of K, namely K2 and K3 stand for the
V̇1 = −λ1 (u3 , u2 )[V1 − V1∞ (u3 , u2 )] (1) curvatures of the beam in the two planes (G, t1 , t3 )(t, X)
V̇2 = 0 (2) and (G, t1 , t2 )(t, X) (see figure 3) while the first component
V̇3 = −λ3 (uq , V1 )[V3 − V3∞ (uq , V1 )] (3) K1 stands for the torsion strain field. G(X) the center of
mass of the X section.
w˙p = −λp (up , V1 )[wp − wp∞ (up , V1 )] (4) As mentioned previously, the 3D control of the eel-
q˙1 = −λ1q (uq , V1 )q1 − λ2q (uq , V1 )uq (5) like robot is realized here by 3D robot’s body movements
w˙q = (1 − 2µq (uq , V1 )|q1 |)q˙1 (6) without using its pectoral fins. For more informations about
the without pectoral fins swimming feasibility problem, the
w˙r = ρV̇1 − λr V1 (ρ − u1 ) (7)
reader can refer to [7].
ρ̇= −λr (ρ − u1 ) (8) The robot’s back part can be defined by (See figure 3) :
1
Q̇ = M (w)Q (9) χback = [Xb , L] (11)
2
where Xb is a given material abscissa and L is the robot’s where for all variable F , F (k) is a short notation for F (kτs )
length. and δF = F (k + 1) − F (k).
Recall that λ1 , V1∞ , λ3 , V3∞ , λp , p∞ , λ1q , λ2q , µq are the
identified parameters as functions of the control vector u
[see (15)]. Let :

T
δ = δV1 δV3 δwp δwq δwr (21)
G(X)
T
= δ1 δ2 δ3 δ4 δ5 (22)

δ ∈ ∆(τs , V1 , u) = [δmin , δmax ] (23)

Fig. 3. Frames and parametrization. denotes the intermediate unknown vector that has to
be computed in order to achieve the tracking objective.
The control law is based on the following predefined Once these increments are computed, this gives reference
temporal structure of the undulation laws K1 and K2 : values for the lower level control vector u. The later is
then computed by a constrained optimization step. More
2π precisely, having the robot’s velocity VA (k) at a given
∀X ∈ χback , K1 (X, t) = uq (t) · cos( t) (12)
T instant k, the control strategy consists in the computation of
2π
K2 (X, t) = up (t) · cos( t) (13) the desired velocity VAd (k + 1) to reach at the next sampling
T time (k + 1). Once this desired velocity is computed,
where uq ∈ [−umax
q , umax
q ] and up ∈ [−umaxp , umax
p ] are the increment δV that is directly linked to the increment
used as control variables. uq is the twist angle and up is the vector δ [through equations (16)-(17)] enables to derive an
pitch angle. optimization problem in which the roll angle stabilization is
As for the undulation law K3 , the following structure is also taken into account. The whole formalism is precisely
adopted in accordance with biological observations [4] : given in the next sections.
X t
K3 (t, X) := u3 (t) · A(X, u2 (t)) sin( − ) + u1 (t), (14) The δ’s lower and upper bounds (δmin , δmax ) are computed
λ T on line according to the current robot’s state. More precisely,
where u3 ∈ [0, umax
3 ], u2 ∈ {−1, 1}, and u1 ∈ [0, umax 1 ]. at each sampling period, given the robot’s state, the control
The control input u2 ∈ {−1, 1} defines whether the ampli- vector saturation levels (u ∈ [umin ,umax ]) and the static
tude of undulations is bigger at the eel’s tail or the eel’s head. maps of the different model’s parameters (that are dependent
This is used to enhance acceleration or deceleration accord- of the control input), the δ’s lower and upper bounds are
ing to the velocity related control requirements. When the computed using the equations (16)-(20).
undulation law (14) is used with u1 ≡ 0, a strait movement is
asymptotically obtained while constant non vanishing values A. Defining the desired velocity VAd (k + 1)
of u1 asymptotically lead to circular trajectories.
Note that (12)-(13)-(14) define a finite dimensioned Let PA , VA designate the robot’s position and velocity at
parametrization of the control input leading to the control a given instant k and PC the desired objective (see figure
+ +
vector 4). PA0 , VA0 are respectively the position and velocity that
would be obtained at the next sampling instant (k + 1) if
u := (up , uq , u1 , u2 , u3 ) (15) δ(k) = 0 is applied during the sampling period.

The controller has to appropriately modify the control vector

in order to steer the head towards the desired 3D position
as well as to realize the rolling angle stabilisation and the
velocity control.

IV. T HE PROPOSED FEEDBACK

After Linearization and using a small sampling period τs ,
the equations (1), (3), (4), (6) and (7) become :

δV1 = −λ1 τs [V1 − V1∞ ] (16)

δV3 = −λ3 τs [V3 − V3∞ ] (17)
δwp = −λp τs [wp − wp∞ ] (18) Fig. 4. Control strategy

δwq = −(1 − 2µq |q1 |)(λ1q q1 + λ2q uq ) (19)

Note that all quantities are expressed in the earth frame.
δwr = ρτs (V̇1 − λr V1 ) + λr τs V1 u1 (20) The ideal velocity that can realize the robot’s mission is
 C P+ P
directed by EAC , where EAC = kPA0 + C. Rolling angle stabilization related term
A0 PC k
More precisely, the ideal velocity would be given by : The control of the rolling angle amounts to control prol :=
+
! t20 · E3 around 0. This amounts to control the head’s roll
kPA0 PC k q
VId = min Vmax , +
, 2kPA0 PC kγdmax EAC (24) velocity wp suggesting the following ideal relation :
τs

w˙p = −γ1 (wp − wpd ) ; wpd = −γ2 prol

since this takes into account the achievable maximum
velocity Vmax , the sampling nature of the control law and the that can be written after Linearization as follows :
fact that one would like to reach the objective at zero velocity
which imposes some deceleration margin that is compatible δwp = −γ1 τs (wp + γ2 prol )
with the maximum deceleration module γdmax . γ1 and γ2 are control parameters.
Now the ideal velocity is generally not achievable in Here again, using the incremental equations (18)-(20), the
within the actuator constraint, that is why an interpolation is above can be written in terms of the incremental vector δ in
introduced through the parameter λ leading to the following the following compact form :
desired velocity :
A1 · δ = B1 (30)
+ +
VAd (k + 1) = VA0 + λ(VId − VA0 ) (25)
where :
where λ ∈ [0, 1] is a parameter that is adapted on line


A1 = 0 0 1 0 0 (31)
according to the current configuration in order to tackle B1 (k) = −γ1 τs (wp (k) + γ2 prol (k)) (32)
dynamically the actuator saturations. In this paper however,
this parameter is determined through worst case calibration D. Controller formulation
for simplicity. Gathering together the two requirements, the following
constrained optimization problem can be derived for the
B. Velocity tracking related term computation of δ :
The constrained optimization problem leading to the com- min (η · δ T δ + α1 kAδ − Bk2 + α2 kA1 δ − B1 k2 ) (33)
δ∈∆
putation of δ is obtained via constrained trade-off between
the velocity tracking consideration and the roll angle stabi-
lization concern. In this section, the velocity tracking related δ ∈ ∆(τs , V1 , u) = [δmin , δmax ]
term used in the optimization problem is derived. The robot where :
velocity in the earth frame at an instant k is given by the • δ is the increment to be computed and u is the current
equation (10) : control value. [see (15)]
  • η, α1 , α2 are control parameters.
V1 (k)
• τs is the sampling period.
VA (k) = Rq (Q(k))  0(k)  (26)
V3 (k) The first part of the equation (33) is a regulation term,
the second part allows to take into account the velocity
Simple computations show that as long as the linear tracking while the third part accounts for the rolling angle
approximation is used, the conditions of perfect tracking stabilisation.
VA (k + 1) = VAd (k + 1)) is equivalent to the following linear This optimization problem can be written as a time varying
equation in the unknown vector δ : quadratic problem :
1
A·δ = B (27) min ( δ T Sδ + f T δ) (34)
δ∈∆ 2
where : with the constraints :
δmin ≤ δ ≤ δmax
A = A(Q(k), VAd (k + 1)) (28)
B = B(Q(k), VAd (k + 1), V0 (k)) (29) where :
S = 2(α1 AT A + ηE + α2 AT1 A1 )
in witch A is a (3 × 5) matrix, B is a (3 × 1) matrix and
f = −2(α1 AT B + α2 AT1 B1 )
 
V1 (k)
V. S IMULATIONS
V0 (k) =  V2 (k) 
V3 (k) In this section, some numerical simulations are proposed
to assess the efficiency and underline some interesting fea-
is the robot velocity in the head mobile frame [see (1), (2) tures of the proposed solution, using the reduced and the
and (3)]. complete continuous model.
A. The robot parameters Figure 5 shows the behavior of the head’s coordinates as
The exhaustive definition of the model parameters is given well as the evolution of the rolling angle arcsin(prol ) for two
in [3]. Let us mention here that the length of the robot different control input bounds [u := (up , uq , u1 , u2 , u3 )] :
is L = 2.08 m and all the cross sections are ellipsoidal umin = (−4◦ , −20◦ , 0, −1, 0) (36)
with evolutive dimension that reproduces a quite realistic umax = (4◦ , 20◦ , 0.5, 1, 1.4) (37)
and faithful form (the tail is thinner than the central body).
or
B. Control related parameters
umin = (−4◦ , −10◦ , 0, −1, 0) (38)
• The sampling period τs = 1.2 s.
umax = (4◦ , 10◦ , 0.5, 1, 0.8) (39)
• 35 % of the robot’s length are used as the robot’s
back part that realize the pitch and the twist movements The evolution of the control input (up , uq , u1 , u2 , u3 )
(swimming without pectoral fins)[see (11)]. during these scenarios is depicted on figure 6.
• The control input u3 ∈ [0, 1.4] and u2 = [−1, 1] (see
section III) : For u3 = 1.4, we obtained a maximal 2
1.5 u3
velocity of 50cm/sec (u3 = 1) in acceleration mode 1
(u2 = −1). 0.5
• The twist angle uq ∈ [−20◦ , 20◦ ] (see section III). 0
0 10 20 30 40 50 60 70 80 90 100 110 120 130
• The pitch angle up ∈ [−4◦ , 4◦ ] (see section III). 2
u2
1
• The maximal body curvature umax 1 = 0.5. 0
(u1 ∈ [0, umax
1 ]), see section III. -1
-2
0 10 20 30 40 50 60 70 80 90 100 110 120 130
• λr = 1.0 [see (8)]. 1
u1
• λ = 0.5 [see (25)]. 0.5
• γ1 = 1, γ2 = 1 [see (30)].
0
• α1 = 1, α2 = 10−2 [see (33)]. 0 10 20 30 40 50 60 70 80 90 100 110 120 130
8
• η = 10−10 kAT Ak + 10−6 [see (33)]. 4
up
• Vmax = 40 cm/sec [see (24)]. 0
-4
-8
C. Manoeuvre description 0 10 20 30 40 50 60 70 80 90 100 110 120 130
30
20 uq
Two set-point changes are successively and simultaneously 10
0
done on the three coordinates of the desired position Pd . The -10
-20
robot is initially at rest. The desired state is then defined by -30
0 10 20 30 40 50 60 70 80 90 100 110 120 130
the following expression : Time (Sec)
Fig. 6. Evolutions of the control input u during the scenarios of figure 5.
The solid and squared lines correspond to the constraints (36), (37) and
Evolution of the (x, y) head coordinates (m) (38), (39) respectively
15
10
5 The evolution of the robot velocity in the earth frame
0
during the same scenarios is presented in figure 7.
-5
0 10 20 30 40 50 60 70 80 90 100 110 120 130 Figures 8 and 9 show the evolution of the different
Evolution of the (z) head coordinates (m) variables in a trajectory tracking scenario tested on the
1.5
1
0.5
continuous complete model [3]. The desired trajectory is
0 geometrically defined as shown in figure 8.
-0.5
-1
-1.5
0 10 20 30 40 50 60 70 80 90 100 110 120 130
VI. C ONCLUSION
Evolution of the rolling angle arcsin(prol ) (deg) In this paper, a complete control scheme for 3D movement
20
15
10 of an eel-like robot is proposed. A multi-variable feedback
5
0
-5
design that enables the tracking of a desired 3D position of
-10
-15 the Eel head as well as the stabilization of the rolling angle is
-20
0 10 20 30 40 50 60 70 80 90 100 110 120 130 presented. This is done using the reduced model that enabled
Time (seconds) the derivation of a constrained quadratic optimization prob-
Fig. 5. Behavior of the controlled robot under the two successive set-point
changes given by (35). Solid and squared lines represent the behavior for lem that can be solved on-line in order to better achieve the
two different control input bounds [see (37) and (39)]. trajectory tracking as well as the roll angle stabilization tasks.
The controller is tested using the reduced and the continuous
(
T model [3] for many 3D scenarios. Future work concerns the
6, −3, 1 for t ≤ 60 s implementation on the prototype (under construction) as well
Pd (t) =
T (35)
14, −5, −1 for t > 60 s as the explicit handling of the actuator saturation in terms of
 Evolution of VA1 (m/s) [see (26)] [15] M. S. Triantafyllou, G. S. Triantafyllou, and R. Gropalkrishnan.
0.5 Optimal thrust development in oscillating foils with application to fish
0.4
propulsion. J. Fluids Structures, 7:205–224, 1993.
0.3
0.2
[16] P. A. Vela, K. A. Morgansen, and J. W. Burdick. Underwater
0.1
locomotion from oscillatory shape deformations. In Proceedings of
0 the 41st IEEE Conference on Decision and Control, pages 2074–2080,
0 10 20 30 40 50 60 70 80 90 100 110 120 130
Las Vegas, 2002.
Evolution of VA2 (m/s) [17] J. Yu, L. Wang, and M. Tan. A framework for biomimetic robot fish’s
0.2

0.1 design and its realisation. In American Control Conference, pages

0
1593–1598, Portland, 2005.
-0.1

-0.2
0 10 20 30 40 50 60 70 80 90 100 110 120 130
Evolution of VA3 (m/s)
0.2

0.1
6
0
5
-0.1

-0.2 4
0 10 20 30 40 50 60 70 80 90 100 110 120 130
Time (seconds) 3

Z
Fig. 7. Evolutions of the robot velocity components in the earth frame 2
during the scenarios of figure 5.
1

torque rather than in terms of amplitudes of undulation as it -1

15
is done in the current paper. 10
6
8

5 4
2
R EFERENCES 0 -2
0
-4
-5 -6
[1] M. Alamir, M. El Rafei, G. Hafidi, N. marchand, M. Porez, and Y X
F. Boyer. Feedback design for 3d movement of an eel-like robot. In
Proceedings of the IEEE Int. Conf. Robotics and Automation, pages Fig. 8. The 3D trajectory of the eel’s head during a 3D trajectory tracking
256–261, Roma, 2007.
[2] P. R. Bandyopadhyay. Trends in biorobotic autonomous undersea
vehicles. IEEE Journal of Oceanic Engineering, 30(1):109–139,
January 2005.
[3] F. Boyer, M. Porez, and W. Khalil. Macro-continuous computed torque Evolution of the (x, y) head coordinates (m)
15
algorithm for a three-dimensional eel-like robot. IEEE Transaction on 10
Robotics and Automation, 22(4):763–775, August 2006. 5
[4] J. Carling, T. L. Williams, and G. Bowtell. Self-propelled anguilli-
0
form swimming: simultaneous solution of the two-dimensional navier-
-5
stokes equations and newtons laws of motion. Journal of experimental 0 20 40 60 80 100 120 140
biology, 201:3143–3166, 1998. Evolution of the (z) head coordinates (m)
6
[5] Jack C. K. Chou. Quaternion kinematic and dynamic differential
4
equations. IEEE Transactions on Robotics and Automation, 8(1):53–
2
64, February 1992.
[6] J. E. Colgate and K. M. Lynch. Mechanics and control of swimming: 0

A review. IEEE Journal of Oceanic Engineering, 29(3):660–73, July -2

0 20 40 60 80 100 120 140
2004. Evolution of the rolling angle arcsin(prol ) (deg)
[7] M. El Rafei, M. Alamir, N. Marchand, M. Porez, and F. Boyer. Motion 90
60
control of a three-dimensional eel-like robot without pectoral fins. In 30
Proceedings of IFAC World congress, Seoul, 2008. 0
[8] R. Mason and J. W. Burdick. Experiments in carangiform robotic -30
-60
fish locomotion. In Proceedings of the IEEE Int. Conf. Robotics and -90
0 20 40 60 80 100 120 140
Automation, pages 428–435, San Francisco, 2000.
[9] K. A. McIsaac and J. P. Ostrowski. A geometric approach to Time (seconds)
anguiliform locomotion modelling of an underwater eel robot. In Fig. 9. Behavior of the controlled robot in the scenario of figure 8.
Proceedings of the IEEE Int. Conf. Robotics and Automation, pages
2843–2848, Detroit, 1999.
[10] K. A. Morgansen, V. Duidam, R. J. Mason, J. W. Burdick, and R.M.
Murray. Nonlinear control methods for planar carangiform robot fish
locomotion. volume 1, pages 427–434, 2001.
[11] K. A. Morgansen, P. A. Vela, and J. W. Burdick. Trajectory stabiliza-
tion for a planar carangiform robot fish. In Proceedings of the IEEE
Int. Conf. Robotics and Automation, pages 756–762, Washington,
2002.
[12] M. Sfakiotakis, D. M. Lane, and B. C. Davies. Review of fish
swimming modes for aquatic locomotion. IEEE Journal of Oceanic
Engineering, 24(2):237–252, April 1999.
[13] J. C. Simo. A finite strain beam formulation. the three dimensional
dynamic problem. part i: formulation and optimal parametrization.
Comp. Meth. Appl. Mech. Eng., 72:276–304, 1989.
[14] M. S. Triantafyllou and G. S. Triantafyllou. An efficient swimming
machine. Scientific American, 272:64–70, March 1995.