Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

ADEQUATE SIX DEGREE OF FREEDOM CONTROL OF ASTABLE

UNMANNED UNDERWATER VEHICLES

'Dept. of Elec. and Comp. Eng., American University of Beirut. P.O. BOX11-0236 Riad El-Solh? Beirut 1107 2020;
Lebanon, Fax: 961-1-744462, email fuad@aub.edu.lb
'Dept. of Elec. and Comp. Eng., American University of Beirut, P.O. Box 113-5701 Hamra, Beirut, Lebanon, email:
antoine@ieee.org

Absnoct - Astable unmanned underwater vehicles 11. ASTABLE uuv CONTROL


(UUVs) are excellent in underwater maneuverability but
challenging in terms of control. An existing Lyapunov based A . Introduction
adaptive nonlinear vehicle controller is enhanced by
replacing its adaptation law with composite adaptation W V s in general are neutrally buoyant rigid bodies.
common in robotics. A novel online trajectov scaling (In practice, slight positive buoyancy is sustained so that
algorithm is developed to prevent actuator saturation and a the vehicle can passively float to the surface in case of a
Matlab based graphical simulator is designed to be a test fault). In particular, astable UWs have coincident
bed on which control performance i s evaluated. centers of gravity and buoyancy. To specify the vehicle
position and attitude, two reference frames are
1. INTRODUCTION considered: a body fixed reference frame called the B-
To achieve good underwater maneuverability, astable frame and an earth fixed (inertial) reference frame called
unmanned underwater vehicles are intentionally designed the I-frame. The z-axis of the 1-frame points upwards
to have their centers of gravity and buoyancy coincident along the vertical. The choice of the x- and y-axes is of
or very close together. This results in attitude instability no concem in simulation studies as long as the right hand
in roll and pitch besides the instability in yaw and vehicle rule is respected. For simplicity, the origin of the B-
position that already exists in stable and astable vehicles. frame is taken to coincide with the centers of gravity and
Countering instability necessitates the use of a control buoyancy and its axes parallel to the vehicle principal
system that can actively regulate vehicle motion. axes, Fig. 1 shows a possible U W shape and the
However, a mere regulating controller is not sufficient for appropriate B-frame.
most applications. In the case of autonomous underwater
vehicles ( A W s ) , a trajectory is generated according to
the mission parameters specified by the user. This
requires a tracking controller that can regulate the vehicle 0 "

about a trajectory not just a point.


A globally convergent Lyapunov based parameter
adaptive tracking controller was devised by Slotine and
Li [I] that applies to robots and all other dynamic systems
modeled as linearly parameterizahle. UUV tracking
control is much more involved, however, due to the fact _..' X X

that, unlike robots, kinematics get into the control


equations. A globally stable 6 degree-of-freedom (DOF)
Fip. 1. .4 rypical UUV body frame.
Lyapunov based position and attitude tracking controller
was derived by Fossen and Fjellstad [2] using the
Vehicle position is expressed by the set of coordinates
singularity free 4 parameter Quatemion to express
of the center of the B-frame referred to the I-frame.
attimde. The controller was then extended into a
Vehicle attitude is expressed by the matrix expressing
parameter adaptive version [3] following the general
the rotation of the B-frame with respect to the I-frame.
scheme of [I].
This representation, however, is redundant and
The adaptation scheme used in [ l ] and [3] is constant
impractical for all purposes. Only 3 parameters are
gain tracking error based adaptation. Instead, we propose
required to represent attitude, however, minimum
the use of Bounded-Gain-Forgetting composite
parameter representations are singular at certain points
adaptation devised by Slotine [4] for robots as the
and their use induces control instabilities. The
adaptation scheme. We show the superior performance of
Quatemion is a singularity free 4 parameter
the proposed scheme in the following two sections. In
addition, a novel online trajectory scaling algorithm was representation of attitude. If the position is X E &,,
developed to deal with the problem of actuator saturation. and the attitude is q €Rdx,then vehicle pose is

0-7803-7369-3/02/$17.00 Q 2002 IEEE


430

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:11 UTC from IEEE Xplore. Restrictions apply.
Due to the conditions mentioned in section II.A,
<= E R,xl.The tracking problem is to bring c ( t ) g ( q )= 0 . Knowing that the model represented by ( I )
is linearly parameterizable, the bracketed expression in
to <,(t) as fast as possible, where <,!(t) is the desired ( 5 ) can be linearly parameterized.as a regressor form:
trajectory. The controller is expected to work in a Y(u,ur,v,)a=MVr +C(u)u, +D(u)u,. (6)
changing environment, so it has to be adaptive.
where a ER,,,,is the parameter vector and
B. Constant gain fi-ackingEwer Based fTEB) Y ( u ,u r ,Ur) E If$,,,, is the known regressor matrix (see
adaptation
[ 5 ] for an expression of this matrix). Let a he the
UUV motion is fully described by two vector
differential equations: dynamics and kinematics. estimate of
.
a , and
.
a La-a be the estimation error.
Dynamics is expressed as [3]: Then, i = - a . Assuming that vehicle parameters are
Mi,+C(u)u+D(u)u+g(q)=r (1) unknown but varying very slowly with time (i.e. U = U),
where M E R,,, is the symmetric positive definite then U = 6.A positive definite Lyapunov function
inertia matrix. including hydrodynamic added mass, candidate is suggested:
C(u)E R,,, is the skewsymmetric Coriolis and 1
V= -[
2
S'MS + C T ' r ' C i ] >0, VS # 0 and
centrifugal matrix. D(u)E I&6x6
is the positive definite
hydrodynamic drag matrix, g ( q )E R651
is a vector of
Vi#O (7)
where TIalsis a constant symmetric positive definite
gravitational and buoyant forces and moments, r6x1
is
adaptation gain matrix. Taking the time derivative of
the vector of applied forces and moments, and both sides of (7):
u=[vT w']' where v ER),^ and are V = S' ( r - Y ~ ) - S ' D ( U ) S +irr-li (8)
respectively the linear and angular velocities of the B- Substituting the adaptation law:
frame with respect to the I-frame expressed in the B-
a =- r y r s (9)
frame. (See [3] for expressions of M , c(u),D ( v ) , and the control law:
and g ( q ) .) r = Ya - K,s (10)
Kinematics is exoressed as 131: we get:
V = -sT [ K D+ D ( u ) ] s(negative semi-definite)
This guarantees that V will keep decreasing until S ,
where R ( q ) is the rotation matrix whose Quatemion and thus V settles at zero. V will settle at some
representation is q and constant positive value, therefore, parameter error is only
guaranteed to converge to a constant value, which may
be zero under some conditions.
(3)
Although tracking errors eventually converge to zero,
the parameter update law largely affects system
7 and EE R3xlare the components of q defined as perfonnance during the transient period, mainly, the

q =[I]. 13x3is the 3 x 3 identity matrix, and


transient behavior of the tracking errors and the
requested forces and torques. As will he seen in the
following section, Bounded-Gain-Forgetting composite
S ( & )E R,,, is defined as: adaptation yields faster convergence of parameter
estimates and, thus, a better behaved system
performance.
S(E)=S([EI E, .;Ir)= C. Boanded-Ga;17-For.tting (BGF) composite
adaptation
A globally stable 6 DOF controller is given by [3]: In composite adaptation, both tracking and prediction
r = MVr+C(u)u,+D(u)u,.+ g ( q ) - K , s errors are used to update the parameter vector, so the
(5)
counterpart of (9) in composite adaptation is formed by
where K , ER,,, is a constant positive definite adding a new prediction error dependent term.
matrix, U, ER,, is the virtual velocity vector, and Additionally, in variable gain composite adaptation laws,
the constant gain matrix is replaced by a positive
s = U - U,.is the virtual body fixed velocity error vector. definite time-varying matrix P ( t ) . Composite
(See [3] for details.)

43 1

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:11 UTC from IEEE Xplore. Restrictions apply.
adaptation schemes differ by the method of varying the where e, is the prediction error defined as:
adaptation gain p ( t ) [4].
e, = W c i = W u - W a = W & - y (18)
I n TEB adaptation, undesirable effects are induced by
and y is the result of filtering r by the same filter
the dependence of the update rate b on Y(U,UV,U,).
that produced W ( u ) from Yo(u,U). Substituting the
When Y is "small", adaptation becomes very slow.
control law of (IO), the gain update law of (13) and (14)
Alternatively, when Y is "large", effects similar to those
caused by a large adaptation gain r
appear. In BGF
and the parameter update law of (17) in (16), v
becomes:
composite adaptation, p ( t ) is made to decrease when
V=-sr[K,+D(u)]s
the "magnitude" of a matrix W (related to Y ) increases.
When W decreases, p ( t ) tends to forget the "large" 1 -r - 1 -
--,la P a (negative semi-definite)( 19)
W and gradually increases maintaining an upper bound. 2
Define Y , ( U , ~ ) E Was:~ ~ ~ ~ 1-
--a7WTWi s0
Y,(u,U)a = MU+C(v)u+D(u)u (11)
2
(12) and (19) guarantee the eventual convergence to
Y,(u, l j ) depends on d which is hard to obtain. This zero of tracking error s,and prediction error w6 while
problem is solved by parameterizing W ( U ) ,a lowpass TEB adaptation only guarantees that of s. Because v
filtered version of Y o ( u , b ) ,instead of Y,(u,U) itself is only negative semi-definite, V settles at a constant
[4]. (See [ 5 ] fordetails.) value which can be zero under some persistency of
A positive definite Lyapunov function candidate is excitation conditions [4]. Only then can parameter
suggested: estiination errors settle at zero. Exact parameter
estimation, however, is not necessary to achieve zero
V = -1s T M s + -1a-r P - I ( t ) i (12) tracking errors and composite adaptation still performs
2 2 better than TEB adaptation even in the absence of
P(1) is a positive definite adaptation gain matrix. If persistency of excitation.
P ( 0 ) was chosen to he symmetrical, it will continue to
111. SlMULATlOS STUDY
he so for all I because it is updated according to the
following rule: A . 1nrroductioii
d To test the enhancements introduced by the use of
--P-'(r)=-A(r)P-l(t)+W'W (13)
dt BGF composite instead of constant gain TEB adaptation,
with A ( t ) being a variable forgetting factor defined a Matlab based graphical simulator was built. Dynamics,
kinematics, and control are implemented using Simulink
as:
blocks. After the model is run, results are fed into Matlab
AV)= 4(1-llpll/~o) (14) and an animation equivalent to these results is generated.
The graphical simulator is, therefore, divided into two
and l/PII is the norm of P . The nonil of a matrix has functional parts: a simulator implemented in Simulink
several possible definitions [6]. The one used in this blocks, and an animator implemented in Matlab code.
paper is the Frobenius n o m defined by: The two loops run separately to achieve larger frame
rates. The simulator window is shown in Fig. 2.
(15)
",..
The two constants k,, > 0 and > 0 in (14) denote
respectively the upper bound of the gain matrix nonn and
the maximum forgetting rate.
Taking the time derivative of both sides of (12):
-7 .I: 1 -r d -
V = s r M i + a P a+-a -(P )a
2 dt
1 -r d
)6
= sr [5 - Ya]- srD(u)s+ i r P - ' d + - a -(P-'
2 dt Fig 2 Simulatorwindou
(16)
Consider the adaptation law.
6(t)=-~(t)[~~s+W~e,] (17)

432

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:11 UTC from IEEE Xplore. Restrictions apply.
B. Trajectory description

The desired trajectory is fd(f) = [zi:;] with


analytical integration
expressions.
to

~~~
obtain the remaining

q,(t) = [ 7ld

$I
(0
(1)
1. A trajectory continuous in its first and

second derivatives was used as the reference trajectory


that the UUV is required to track. Its spiral shape makes it
suitable for graphical simulation and its smooth velocity
profile makes it practical to implement in terms of
actuator demand. This trajectory is given by the following F i i . 4. A 3 D trace ofTG7
equations:
x ‘ , ( ~ ) = x , , , c o s ( c ~ t ) - ~ : , ( t ) = ~ ~ , z,,(t)=v2t
,,sinwt,
C. Simaiatior~
4,,(?)=
wt+n/2,~,(/)=cos-,and
4d For the purpose of dynamics simulation. the true value
2 of the parameter vector will be taken to be [3]:
4’i
c d ( t )= sin(-)n(,
2
where x,,, and y,<, are constants. 0 and v, are

functions of time to be specified next, and


constant vector. To guarantee continuous
nd is a
<,,, f<,, and
U =col
1 215,265,265,40,80,80,
70,100,100,30,50,50,
100,200,200,50,100,100
(See [ 5 ] for details on obtaining
D ( v ) from this vector.)
M,
(20)

c ( v ) , and

cd,w , h ,# , I>:, C:, and V: must all be continuous. The initial estimate i ( 0 ) of the parameter vector will
The scheme of Fig. 3 satisfies this requirement. It can be be set to zero. In TEB adaptation, the gain matrix has
used for both # and ijr when the constants a and b to be specified. A very large r
is inadequate for both
digital implementation and computer simulation. A very
(defined in the figure) are set to appropriate values.
small , on the other hand. will result in a very slou,
adaptation. For the purposes of comparison with the

” I BGF composite adaptation law, and given the chosen


step size, an acceptable
This value is r
is found by trial and error.
= 50. In BGF composite adaptation,
the values that will be used for the adaptation parameters
are To=o.OO1~,,xl,, 4 = l o , and k,, =10000.
Fig. 5 contains the simulation results. Comparing the
solid with the dashed curves in Fig. 5(a) and Fig. 5(b),
we can see that the deviation from the desired trajectory
to be significantly smaller in the BGF case than in the
Fig 3 ~ protile TEB case. Note that the superiority of position and
attitude tracking in the BGF case is much clearer in
Assume the following numerical values graphical simulations than in the figures.
4 Making a similar comparison also in Fig. S(c) and Fig.
xn,= l o , yn,= l o , n, =[0 0 1Ir, a, =-x0.4, 5(d), we can observe that forces and torques in the BGF
3 case are found to he significantly better behaved than in
4 the TEB case. They are smaller in peak magnitude and
a,, = -x0.2, brz = b, = 2 . A 3D trace of the have no violent oscillations and sharp transitions that are
- 3
trajectory obtained for these numerical values is shown in very hard for physical actuators to realize.
Fig. 4. Fig. 5(e) and Fig. 5(f) show 6 sample elements of the
To avoid numerical differentiation, the time derivatives IS-element parameter vector. In comparing dashed with
solid curves, we can see that parameter adaptation in the
of f d require analytical expressions for GJ, h ,&, vI, BGF case is superior to that in the TEB case in that it is
v: , and G z . These expressions can be formed from Fig. 3 faster, oscillation free, and that most parameters
by converting the plot into time equations and using

433

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:11 UTC from IEEE Xplore. Restrictions apply.
I I

Fig. 5 . Constant gain TEB (solid curve) versus BGF composite (dashed curve) adapfation. ( a ) Position X . (b) AttiNde error
-
4 . ( c ) Forces. id) Torques.
( e )Elemenrs I.2. and 3 o f the parameter vector. (0 Elements 10. 11. and I2 of the parameter vector.

converge to their true values, and all of them converge to Parameter adaptation may cause excessive forces or
nearby regions of the true values. torques to be requested.
If the UUV is departed either in position or in
I\'. ONLINE TRAJECTORY SCALING attitude or in both from the reference trajectory by a
The control law represented by ( 5 ) as well as the disturbance, the control algorithm, attempting to
adaptive version represented by (IO) are vulnerable to bring the tracking errors exponentially to zero,
actuator saturation, meaning that they can go into serious requests excessive forces or torques.
instabilities in case a requested force or torque was not We propose a novel online trajectory scaling algorithm
met by available actuators. Saturation can occur in any of to counter actuator saturation when it occurs. The
the following scenarios: following is a brief overview of the algorithm, for details
The specified trajectory may he too fast for the see [ 5 ] . The algorithm keeps slowing the specified
available actuation limits. desired trajectory as long as saturation persists, perhaps
down to a full stop. On the other hand, the reference

434

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:11 UTC from IEEE Xplore. Restrictions apply.
trajectory is made faster when actuation capability is requested control forces and torques. This paper
available. presented the derivation of the proposed UUV adaptive
We first define a scalar saturation margin. Under control scheme. In addition, simulation and animation
saturation, the margin is equal to the scaled element of were conducted to validate certain advantageous
the control vector that has undergone the most serious characteristics of the proposed scheme. The paper finally
saturation. Under no saturation, the margin is equal to the presented a brief overview of a novel online trajectory
scaled element of the control vector that is closest to its scaling algorithm.
respective limit. Depending on the sign of the margin,
one of two filters is selected. One filter speeds the VI. REFERENCES
trajectory up and the other slows it down. Filters are [I] JLJ. Slotine and \I1. Li. "On the adaptive control of robot
designed so that switching from one filter into another manipulators." The Imm~mimmIJoirrnol of Robotics Reseordi.
vol. 6. no. 3. 1987. 49-59.
does not affect the continuity of the original unscaled
[2] 0.-E. Fjellsthd and T. Fassen. "Position and attitude tracking of
trajectory. ALiV's: A Quatemion feedback approach." 1EEE Jounioi qf
The algorithm was tested with the spiral trajectory ~ , 19. no. 4. 1994. 512-518.
Ocronic E n ~ ; m e r i nvol.
described in 1II.B which required a maximum force of [i] O..E. Fiellsrad and T. Fossen. "Singulariry-free racking of
unmanned undenvater vehicles in 6 DOF." m Procredrnps ofdie
6000N along the x-axis of the body frame (Fig. 6). The 37"lEEE ConTerence 017 Decision arid Conrrol. 1994.
saturation limit for this component of the force vector 141 J.-J. Slormc. "Composite adaptive conirol d r a b a t manlpulators."
A,,raniarico. vol. 25. no.4. 1989.509-519.
was set to 500N, which is 12 times less. With initial
[j] A . Majduiani. "The conml of undenvater remotely operated
parameter estimates set to 114 of the correct value, and vehicles." h405,er- of €ngimer;ns T1ws;s. A n w i c o s L l ? r ; w r $ io/
~~
BeirMi.
BGF composite adaptation used in the update. simulation E. Kreysrig. ,Advanced Engineevmx Alorlwn~n!,cs.Wiley. New
[6]
showed that the requested force was slightly above 500N York: 1993.
(Fig. 6). Note that only 4 is considered because of E,
the unscaled trajectory is the largest of all other forces
and torques with respect to its saturation limit and 6 of
the scaled trajectory is the only force that causes a slight
saturation.
The proposed algorithm suggests a method for online
scaling of desired trajectories. It does not guarantee that
all forces and torques will stay within saturation limits.
What it does, however, is that it offers a practical way of
implementing control algorithms that are not designed to
work under actuator constraints. Using such algorithms
in the presence of actuator saturation would result in
perfonnance degradation and sometimes, instability.

-
I_

Yr..

m~

.".
.~ ~..~~~
* , %I ' *I
,,-,"~, , 3% I .$ I

Fig. 6. Effect o f online rajectory scaling on the requested Fx. The solid
E U N ~is Fx needed by the unscaled trajectoly. The dsrhcd curve is Fx
requested under tmjecroly scaling (maximum= 51 OX]. The horizontal
line is the adopted saturation limit (500A').

v. coNcLusloN
We have replaced constant-gain tracking error based
adaptation used in adaptive UUV control, with Bounded-
Gain-Forgetting composite adaptation common in
robotics. The proposed UUV adaptive control scheme
yields faster dynamic parameter convergence, faster
tracking error convergence, and more permissible

435

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:11 UTC from IEEE Xplore. Restrictions apply.

You might also like