ROBUST ADAPTIVE CONTROL OF UNDERWATER VEHICLES: A COMPARATIVE STUDY

Thor I. Fossen and Ola-Erik Fjellstad

University of Trondheim, The Norwegian Institute of Technology, Department of Engineering Cybernetics, N-7034 Trondheim, NORWAY (E-mail:tif@itk.unit.no)
measurement noise. The performance of the adaptive control laws of Sadegh and Horowitz (1990) and Slotine and Benedetto (1990) are compared. Both these schemes require that all states are measured, that is the velocities and positions in surge, sway, heave, roll, pitch and yaw. However, for underwater vehicles it is di cult to measure the linear velocities whereas angular velocity measurements can be obtained by using a 3-axes angular rate sensor. This problem is adressed by designing a nonlinear observer for linear velocity state estimation. The proposed observer requires that the position and the attitude are measured, e.g. by using a hydroacoustic positioning system for linear positions, two gyros for roll and pitch and a compass for yaw. In addition angular rate measurements will be assumed available from a 3-axes rate sensor or a state estimator. It is also assumed that the measurement rate is limited to 2 Hz for all the sensors. Simulation studies with a 3 DOF AUV model are used to demonstrate the convergence and robustness of the adaptive control laws and the velocity state observer. Key Words. ROV, AUV, adaptive control, nonlinear velocity observer, marine systems.

Abstract. Robust adaptive control of underwater vehicles in 6 DOF is analyzed in the context of

1. INTRODUCTION The implementation of advanced adaptive control laws for underwater vehicles in 6 degrees of freedom (DOF) is mainly limited by the performance of the underwater navigation and sensor systems. The surge, sway and heave position can be measured with fairly accuracy by applying a hydroacoustic short or long base-line system whereas attitude usually is measured by 2 gyros (or inclinometers) and a compass. The bandwidth limitation of the control system is mainly due to the hydroacoustic positioning system which usually operates in the interval of 0.1{1.0 Hz. This paper discusses the performance of two stateof-the-art adaptive control laws for robot manipulators which are modi ed to control an underwater vehicle in 6 DOF. In addition, a globally asymptotically stable velocity observer using linear acceleration and position measurements is suggested for estimation of linear velocity. 2. UNDERWATER VEHICLE MODEL The following 6 DOF autonomous underwater vehicle (AUV) equations of motion are considered, Fossen (1994):

where = x; y; z; ; ; ]T is the earth- xed position and attitude vector, = u; v; w; p; q; r]T is a vector of body- xed linear and angular velocities and = 1 ; :::; 6 ]T is a vector of control inputs (force and moment). The matrices and vectors are given in Appendix A. 3. A GLOBALLY ASYMPTOTICALLY STABLE VELOCITY OBSERVER For underwater vehicles it is di cult to measure the linear velocity with good accuracy. A globally asymptotically stable observer can be designed for this purpose. Let:

v !

= u; v; w]T = p; q; r]T

x q

= x; y; z ]T = ; ; ]T

(3)

In some cases it is advantageous to use 3 linear accelerometers to improve the convergence of the observer. The location of the linear accelerometer is given by the coordinates:

r a = x a ; y a ; z a ]T

(4)

3.1. Estimation of Linear Velocity The output from the linear accelerometers are:

M _ + C(

) + D( ) + g( ) = _ = J( )

(1) (2)

_ +! v+ = v _ ra + ! (! !

ra) + g1 (q)

(5) 1

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

x K1
_ ^ v ^ v

ac

K2 J 1 (q)
^ _ x

^ x

S (!)
Fig. 1. Nonlinear velocity observer using position and linear acceleration measurements.

where g1 (q ) is the acceleration of gravity de ned as; see (58) in Appendix A:

g1 (q) = g s ; ?g c s ; ?g c c

]T

(6)

The body- xed acceleration _ in (5) can be written in terms of the acceleration measurement a according to: _ v _ x = =

= 0; I ]. Since (14) is a linear time-varying system, K 1 and K 2 can be computed by applying the Kalman lter algorithm, Gelb et al. (1988): K (t) = P (t)H T R?1 (15) ~ (t) (16) A(t)^(t) + Bu(t) + K (t) x T A(t)P (t) + P (t)A (t) + Q ?K (t)RK T (t) (17) where Q is the process and R is the measurement with _ (t) = ^ _ (t) = P

where ac is an acceleration correction term dened as: _ ra + ! (! ra ) + g1 (q) ac = ! (9) Eqs. (7) and (8) suggest the observer structure: _ = ?S (!) v ~ ^ ^ + a ? ac + K 1 x (10) v _ ~ ^ ^ (11) x = J 1(q) v + K 2 x

S(!) v + a ? ac J 1(q) v
?

(7) (8)

covariance matrices. This require that the timevarying error covariance matrix propagation given by (17) must be computed on-line.

~ =x?x ^ is the position estimation error, where x and K 1 and K 2 are the estimator gain matrices.

3.2.2. Nonlinear Observer Design. An alternative approach to the Kalman lter algorithm is a nonlinear pole-placement algorithm utilizing the structure of the observer dynamics. Notice that: _ 1 (q ) = J 1 (q) S (!) J (18) Hence (10) and (11) can be written as: _ (t) + J 1 (q)K 1 x ~ (t) = 0 ~ ~ (t) + K 2 x x (19) (20) (21)

3.2. Computation of K1 and K2 ~ =v?v ^ and x ~ =x?x ^. Let v 3.2.1. Kalman Filter Approach. Equations (10) and (11) can be written in state-space form: _ ^ ^ S(!) 0 v v = ? (12) _ ^ J 1(q) 0 x ^ x K1 ~ + I (13) 0 u + K2 y ~ (t) (14) A(t) ^ + Bu(t) + K (t) y with obvious de nitions of A(t); B and K , and ^T ; x ^ T ]T and y ~ = H ( ? ^) with u = a ? ac; ^ = v 2 _ (t) = ^
m

An intuitive solution is to choose K 1 and K 2 as:

K1 = J T 1 (q) K 10 ;

K 2 = K 20

which reduces (19) to: _ (t) + K 10 x ~ (t) = 0 ~ ~ (t) + K 20 x x

Eq. (21) satis es the Lyapunov function candidate: 1 (x _Tx _ +x ~) ~ ~ ~ T K 10 x (22) V = 2 + (23) T _ 0 _ = ?x _ K 20 x ~ ~ (24) V

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

pos. etimation errors 3 2 0.04

vel. estimation errors

0 1 2 0 100 200 s 300 400

m/s
0 0.02 0

0.02

100

200 s

300

400

estimated surge and sway pos. 80 60 1.5 1

estimated surge and sway vel.

ypos

40 20 0 0

m/s
100 200 xpos 300 400

0.5 0 0.5 0

100

200 s

300

400

Fig. 2. Position and velocity estimation errors (upper plots) and position and velocity state estimates (lower plots).

Application of LaSalle's invariant set theorem _ ! 0 and x ~ ~ ! 0 in nite time. then shows that x 3.3. Implementation Issues Implementation of (10) and (11) require that x; q; ! and a are measured. If acceleration a is inconvient to measure the observer structure in Figure 1 can be modi ed to use an estimate of a generated by a lter, e.g. (see Figure 3): _ = ? 1 a + K0x ~ ; K 0 = 1 K 00 > 0 (25) a T T

B was used to generate the acceleartion measurements in surge and sway while the heave acceleration was set to zero. The heading of the vehicle was chosen as: (t) = 0:5 sin(0:03 t); r(t) = 0:015 cos(0:03 t) The estimator gain matrices were chosen as The sampling frequency was 2Hz . It is seen from Fig. 2 that accurate estimates of u and v can be generated from noisy position measurements (approx. amplitude of the white noise is 1.0 m) .

K 10 = 0:001I and K 20 = 1:0I .

where T > 0 is the low-pass lter time constant re ecting the vehicle dynamics. ~ x

K0 a
1 T
Fig. 3. Estimation of acceleration a in Fig. 1. K0 is a new lter gain matrix to be computed together with K1 and K2 in Fig 1.

4. ADAPTIVE CONTROL The stability proof and representation of the two control laws to be presented are based on the parameterization: _ + C ( ) r + D( )r = Y (r _ ; r; ) Mr (27) The gravity and buoyancy term g( ) (without loss of generality) is assumed to be perfectly known. In _ ; r; ) is the regressor mathis representation Y (r ~=
? d

trix (known) and is a constant parameter vector (unknown). Let the tracking error be denoted by: (28)

If ! not is measured, ! should be approximated ^ which can be obtained from an angular veby ! _ 0 (slowlylocity observer. Assuming that ! varying attitude) yields: ^ (^ ac ! ! ra) + g1(q) (26)

where d 2 IR6 is assumed to be a smooth reference trajectory. 4.1. The Adaptive Control Law of Slotine and Benedetto (ASB) The adaptive control law of Slotine and Benedetto (1990) with the parametrization of Fossen (1993) is actually intended for 3 DOF spacecraft attitude control. An extension of this work to 6 DOF underwater vehicle control is discussed by Fossen and Sagatun (1991). The ASB adaptive control scheme can be represented in terms of the model 3

3.4. Simulation Study of Observer The nonlinear observer (10) and (11) was simulated with ac = 0 . The AUV model in Appendix

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

Pilot Inputs way-points (x k ,yk ) LOS

c

velocities uc , v c Autopilot

AUV Dynamics

u, v, r Kinematics

x, y

inner servo -loops outer servo-loop (guidance system)

Fig. 4. Way-point guidance by means of a yaw angle autopilot and a surge-sway velocity control system. The line-of-sight (LOS) algorithm is used to generate the desired yaw angle.

parameterization (P), control law (C) and parameter adaptation law (A) according to:

P: M _ r + C ( ) r + D( ) r = Y ( _ r ; r ; ) C: = Y ( _ r ; r ; )^ + g( ) ? J T ( ) K d s _ = ?? Y T ( _ r ; r ; )J ?1 ( ) s A: ^ where K d > 0, ? = ? T > 0 and: _ + ~; s=~ = T >0 (29)

The virtual reference trajectories are computed according to: (30) r = J ?1 ( ) _ r ? 1 ? 1 _ _ r = J ( ) r ? J ( )J ( ) _ r ] (31) where _ r = _ d ? ~ . 4.2. The Adaptive Control Law of Sadegh and Horowitz (DCAL) Straightforward extensions of the robot control scheme of Sadegh and Horowitz (1990) to underwater vehicles yields:

1. Parameter adaption, ASB 2. All parameters known, ASB 3. Parameter adaption, DCAL 4. All parameters known, DCAL The results are presented according to: Section 4.3.1: Perfect measurements Section 4.3.2: Noisy measurements The performance indices are chosen as:
N 1X u(ti ) ? ud (ti )]2 + v(ti ) ? vd (ti )]2 Juv = N i=1 N 1X Jr = N r(ti ) ? rd (ti )]2 i=1

J J

N 1X (ti ) ? d (ti )]2 = N i=1 N 1X T = N (ti ) (ti ) i=1

P: M _ d + C ( d) d + D( d) d = Y ( _ d; d) C: = Y ( _ d; d) ^ + g( d) _ + K f k ~ k2 s] ?J T ( ) K p ~ + K d ~ _ = ?? Y T ( _ d; d )J ?1 ( )s A: ^ where K p = K T p > 0, K d > 0, K f > 0, ? = T ? > 0. The above control law is known as the

where N is the number of samples used in the simulation study. The sampling frequency is chosen as 2 Hz. The DCAL and ASB control laws were tuned according to:

K ASB d K DCAL p K DCAL d K DCAL f

desired compensation adaptive law (DCAL) since the desired states d and d are used instead of the actual states and in C ( d ),D( d ) and g( d).

4.3. AUV Case Study The performance of the ASB and DCAL control laws are studied for the NPS AUV II given in Appendix B. A combined velocity and position control scheme is studied for the following cases: 4

with simultaneously heading and speed control (surge and sway) was considered, see Fig. 4. The desired heading states ( d ; _ d ; d ), surge states (u _ d ; ud ) and sway states (v _ d ; vd ) are computed by using the lters: 2 d = !n 2 f (36) d + 2 ! n _ d + !n Tu _ d + ud = u f (37) Tv _ d + vd = vf (38)

KD (32) KD (33) KD (34) 0 (35) = 0:1I and K D = 5000I . An autopilot for

= = = =

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

Table 1 Performance indices for the ASB and DCAL control laws. Full state feedback and no measurement noise (800 samples).
index ASB DCAL units 1. Adaptive 2. Perfect 3. Adaptive 4. Perfect Model Model 0:0011 3:9892 10?8 0:0011 6:5414 10?6 (m/s)2 0:0003 9:5260 10?6 0:0003 1:3727 10?5 (deg/s)2 0:1249 0:0009 0:1021 0:0015 (deg)2 3:9795 4:5612 3:9753 4:5633 106

Juv Jr J J

Table 2 Performance indices for the ASB, DCAL and PD control laws. Full state feedback and measurement noise (800 samples).
index ASB DCAL units 1. Adaptive 2. Perfect 3. Adaptive 4. Perfect Model Model par. drift 0:0070 0:0067 0:0048 (m/s)2 par. drift 0:0278 0:0212 0:0189 (deg/s)2 par. drift 2:1421 2:4389 2:4539 (deg)2 par. drift 8:9384 8:9091 8:9012 106
2

Juv Jr J J

where the reference inputs f , uf and vf are preltered by a 1st-order lter to avoid acceleration step inputs. The parameters were chosen as = 1:0; !n = 0:1 rad/s and T = 1=!n = 10:0 s. The pre- lters are designed according to:

commanded surge velocity, sway velocity and yaw angle, respectively. The commanded input c can be computed by applying a line-of-sight (LOS) algorithm, e.g. Healey and Lienard (1993):
c (t) = tan?1

Tf u _ f + uf = uc Tf v _ f + vf = vc _ Tf f + f = c where Tf = T , and uc ; vc and c are the

(39) (40) (41)

r d ? ( _ ? _d) The resulting control laws are:

_r = 4

u _r v _r

3 5

=4

ud vd

3 5

(45)

where Y ( ) is given in Appendix B and: _ ASB = ??Y T ( _ r ; r ; )s ^ _ DCAL = ??Y T ( _ d ; d ; d)s ^

ASB = Y ( _ r ; r ; ) ^ ASB ? K D s DCAL = Y ( _ d ; d ; d ) ^ DCAL ? K D s

(46) (47) (48) (49)

yd (k) ? y(t) xd (k) ? x(t)

(42)

where xk and yk (k = 1...N) represents the waypoint coordinates. Care must be taken to select the proper quadrant for c (t). After the quadrant check is performed, the next way point can be selected on a basis of whether the vessel lies within a circle of acceptance: xd (k) ? x(t)]2 + yd(k) ? y(t)]2 2 (43) 0 where 0 is the radius. If this constraint is satised the next way point xd (k +1); yd(k +1)] should be selected. The measure of tracking is: s= u ~; v ~; _~ + ~]T (44) Hence the ASB virtual reference trajectories can be computed as:

where ? = 100I . Notice that J ( ) = I for this particular case since _ = r whereas the other two states are velocity controlled. 4.3.1. Perfect Measurements. The numerical results are shown in Table 1. Fig. 5 shows the parameter estimates and Fig. 6 shows the tracking performance of the two algorithms. Notice that the performance of the two adaptive controllers are almost equal. Also notice that the perfect model control laws ( = true ) yield better performance than their adaptive counterparts. 4.3.2. Noisy Measurements. The simulation study in Section 4.3.1 was repeated by adding white noise to the state measurements of u; v; r and . It turned out the ASB algorithm was quite sensitive to measurement noise. For instance adding a zero-mean white noise signal to the yaw angle measurement: 5

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

M par. estimates DCAL 15000 400 300 10000 200 100 0 0 0 100 200 time (s) 300 400 100 0

D par. estimates DCAL

5000

100

200 time (s)

300

400

M par. estimates ASB 15000 400 300 10000 200 100 0 0 0 100 200 time (s) 300 400 100 0

D par. estimates ASB

5000

100

200 time (s)

300

400

Fig. 5. No measurement noise: Parameter estimates for ASB and DCAL.

desired surge vel. 1.5 0.4 0.2 1

desired yaw rate

deg/s
0.5 0 0 100 200 time (s) 300 400

m/s

0 0.2 0.4 0

100

200 time (s)

300

400

desired sway vel. 0.5 10 5 0

desired yaw angle

deg
100 200 time (s) 300 400

m/s

0 5

0.5 0

10 0

100

200 time (s)

300

400

ASB surge/sway vel. errors 0.1 1 0.5 0.05

ASB yaw angle error

deg
0 0.05 0 100 200 time (s) 300 400

m/s

0 0.5 1 0

100

200 time (s)

300

400

DCAL surge/sway vel. errors 0.15 0.1 0.05 0 0.05 0 1 0.5

DCAL yaw angle error

deg
100 200 time (s) 300 400

m/s

0 0.5 1 0

100

200 time (s)

300

400

Fig. 6. Upper plots show the desired states whereas the lower plots show the tracking errors for the ASB and DCAL (no measurement noise).

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

DCAL M par. estimates 20000 15000 10000 1000 5000 0 5000 0 500 time [s] 1000 0 1000 0 3000 2000

DCAL D par. estimates

500 time [s] ASB D par. estimates

1000

4 x 10 ASB M par. estimates

3000 2000

1 1000 0 0 1 0 1000 0

500 time [s]

1000

500 time [s]

1000

Fig. 7. Measurement noise: Mass matrix parameter estimates for ASB and DCAL. Notice that drifting for the ASB algorithm.
actual and desired surge vel. 2 1.5 1 0.5

and

are

actual and desired sway vel.

m/s

m/s
100 200 time (s) 300 400

1 0.5 0 0.5 0

0 0.5 1 0

100

200 time (s)

300

400

actual and desired yaw rate 1 0.5 20 10

actual and desired yaw angle

deg/s

0 0.5 1 0

deg
100 200 time (s) 300 400

0 10 20 0

100

200 time (s)

300

400

Fig. 8. Measurement noise: Actual and desired states for DCAL.

taminated signals:
m= +

1:0 deg

(50)

gave the parameter estimates shown in Fig. 7. It is seen that the for the ASB algorithm, the parameter estimates of 3 = Iz ? Nr_ and 6 = ?Nr both corresponding to the yaw mode are drifting. It is also seen that the DCAL parameter estimates are bounded even for the noisy case. Parameter drift in the ASB algorithm can be explained by considering the update laws for 3 and 2 occurs. This im6 where the square signal m plies that white noise is squared and integrated up to a non-zero term causing parameter drift. For the DCAL algorithm this is avoided since the product between the desired state d and the measurement m is used instead. The performance of the DCAL algorithm with the highly noise con-

um = u + u ; vm = v + v ; rm = r + r ;
m =

+ ;

j uj j vj j rj j j

0:1 m/s 0:1 m/s 0:1 deg/s 1:0 deg

(51) (52) (53) (54)

are shown in Figure 8. It is shown that the performance is satisfactory also for this case. 4.4. Comments on the ASB and DCAL Algorithms It is well known that the ASB control scheme is sensitive to velocity measurement noise, see Berghuis (1993) for instance. For an underwater vehicle body- xed velocities are usually obtained by model-based state estimation through 7

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

noisy position measurements. This implies that the velocity estimates can be contaminated with a signi cant amount of noise. In such cases the ASB algorithm can go unstable due to drift in the parameter estimates. However, some precautions against parameter drift can be taken by small modi cations of the adaptive scheme e.g. by applying the so-called and e1 modi cation schemes, Narendra and Annaswamy (1987). The case study showed that the DCAL was less sensitive to velocity measurement noise than the ASB. The main reason for this is that the actual position and velocity in the DCAL regressor is replaced by the desired state trajectories. 5. CONCLUSIONS In this paper two direct adaptive control laws, referred to as the adaptive control law of Slotine and Benedetto (ASB), and the desired compensation adaptive control law (DCAL) by Sadegh and Horowitz, have been studied in the context of measurement noise. It is concluded that the DCAL algorithm is the only suited algorithm for real-time AUV applications. A globally asymptotically nonlinear observer for estimation of linear velocity is also presented. The observer is intended for implementation together with the DCAL algorithm. The case study also shows that the DCAL with xed parameters is highly robust for measurement noise. It can also be concluded that direct parameter adaptation should only be used if accurate position measurements are available at high sampling frequency (1{10 Hz). 6. REFERENCES Berghuis, H. (1993). Model-Based Robot Control: From Theory to Practice. PhD thesis. University of Twente, Enschede, The Netherlands. Fossen, T. I. (1993). Comments on "Hamiltonian Adaptive Control of Spacecraft". IEEE Transactions on Automatic Control AC-38(4), 671{ 672. Fossen, T. I. (1994). Guidance and Control of Ocean Vehicles. John Wiley and Sons Ltd. Fossen, T. I. and O. E. Fjellstad (1995). Nonlinear Modelling of Marine Vehicles in 6 Degrees of Freedom. International Journal of Mathematical Modelling of Systems JMMS-1(1). Fossen, T. I. and S. I. Sagatun (1991). Adaptive Control of Nonlinear Systems: A Case Study of Underwater Robotic Systems. Journal of Robotic Systems JRS-8(3), 393{412. Gelb, A., J. F. Kasper, Jr., R. A. Nash, Jr., C. F. Price and A. A. Sutherland, Jr. (1988). Applied 8

Optimal Estimation. MIT Press. Boston, Massachusetts. Healey, A. J. and D. Lienard (1993). Multivariable Sliding Mode Control for Autonomous Diving and Steering of Unmanned Underwater Vehicles. IEEE Journal of Ocean Engineering OE18(3), 327{339. Narendra, K. S. and A. M. Annaswamy (1987). A New Adaptive Law for Robust Adaption Without Persistent Excitation. IEEE Transactions on Automatic Control AC-32(2), 134{145. Sadegh, N. and R. Horowitz (1990). Stability and Robustness Analysis of a Class of Adptive Controllers for Robotic Manipultaors. Int. Journal of Robotics Research 9, 74{94. Sagatun, S. I. and T.I. Fossen (1991). Lagrangian Formulation of Underwater Vehicles' Dynamics. In: Proceedings of the IEEE International Conference on Systems, Man and Cybernetics. Charlottesville, VA. pp. 1029{1034. Slotine, J. J. E. and M. D. Di Benedetto (1990). Hamiltonian Adaptive Control of Spacecraft. IEEE Transactions on Automatic Control AC35(7), 848{852. SNAME (1950). The Society of Naval Architects and Marine Engineers. Nomenclature for Treating the Motion of a Submerged Body Through a Fluid. In: Technical and Research Bulletin No. 1-5.

A. GENERAL 6 DOF AUV MODEL

AUV Dynamics. The 6 6 inertia matrix M included hydrodynamic added mass is written:

M
2

M 11 M 12 M 21 M 22

(55)

Expanding this expression in terms of the SNAME (1950) notation yields:

M=

6 6 4

m ? Xu _ ?Xv _ ?Xw _ ?Xp_ zG ? Xq_ ?myG ? Xr_ ?Xp_ ?mzG ? Yp_ myG ? Zp _ Ix ? K p _ ?Ixy ? Kq_ ?Izx ? Kr_

?Xv _ m ? Yv _ ?Yw _ ?mzG ? Yp_ ?Yq_ mxG ? Yr_ mzG ? Xq_ ?Yq_ ?mxG ? Zq_ ?Ixy ? Kq_ Iy ? Mq_ ?Iyz ? Mr_

?Xw _ ?Yw _ m ? Zw _ myG ? Zp _ ?mxG ? Zq_ ?Zr_ ?myG ? Xr_ 3 (56) mxG ? Yr_ 7 ?Zr_ ?Izx ? Kr_ 7 5 ?Iyz ? Mr_ Iz ? Nr_

Sagatun and Fossen (1991) have shown that the Coriolis and centripetal matrix can be de ned in terms of the inertia matrix elements according to:

C(

)=

03 3 S(M 11 1 + M 12 2) ?S (M 11 1 + M 12 ?S (M 21 1 + M 22

2) 2)

(57)

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)

g(
2 6 6 6 4

where S ( ) is a 3 3 skew-symmetrical matrix operator de ned such that S (a)b = a b for a 2 IR3 and b 2 IR3 ; see Fossen and Fjellstad (1995) for details. The vector term D(v)v 2 IR6 is simply a vector containing dissipative forces and moments due to hydrodynamic damping and lift whereas g( ) 2 IR6 is the gravitational and buoyant forces and moments de ned by: ) = gT 1 ( ); gT 2 ( )]T (W ? B) s ?(W ? B) c s ?(W ? B) c c ?(yG W ? yB B) c c + (zG W ? zB B) c s (zG W ? zB B) s + (xG W ? xB B) c c ?(xG W ? xB B) c s ? (yG W ? yB B) s
3 7 7 7 5

W = 53400 N Iz = 2038 kgm m = 5443:4 kg L = 5:3 m

Cd = 0:0034 = 1000 kg/m3

The dynamic pressure is: q = 0:5 L2 = 14045:0 kg/m whereas

?3 qL Xujuj = ?Cd q ?7:6 10? 2 ?5:5 10?3 qL Yv = ?1:0 10?1 q 3 ?3:4 10 qL Nr = ?1:6 10?2 qL2 1:2 10?3 qL Nv = ?7:4 10?3 qL The matrices are:

(58)

Xu _ Yv_ Nr_ Yr

= = = =

Kinematics. Eq. (2) describes the kinematic equations of motion where J ( ) is a 6 6 block diagonal transformation matrix usually de ned in terms of the Euler angles ; and according to:

M= C( D(

"

m ? Xu 0 0 _ 0 m ? Yv 0 _ 0 0 Iz ? Nr_

)=
"

J(

)=

Here J 1 ( ) is the Euler angle rotation matrix dened as: " c c ?s c + c s s J 1( ) = s c c c + s s s ?s cs s s + c c s # (60) ?c s + s s c
cc

J 1 ( ) 03 3 03 3 J 2 ( )

(m ? Yv_ )v ?(m ? Xu _ )u ?Xujuj juj

0 0 0 0 ?Yv ?Yr ?Nv ?Nr

0 0

0 0

?(m ? Yv _ )v (m ? Xu _ )u

(59) )=

Regression Form. This model can be written in regression form:

_ + C (v )r + D(v )r = Y (r _ ; r; v) Mr with parameter vector: = m ? Xu _ ; m ? Yv _ ; Iz ? Nr_ ; and regressor:

?Xujuj ; ?Yv ; ?Nr ; ?Yr ; ?Nv

and J 2 ( ) is the angular velocity transformation matrix de ned as:

J 2(

"

)=

1 s t 0 c 0 s =c

c t ?s c =c

(61)

Nonlinear Model Properties. In Fossen (1994) it is shown that the nonlinear model (1) and (2) satis es the following model properties: _ = 0 (assuming const. 1. M = M T > 0 and M added mass for a deeply submerged vehicle). 2. C ( ) = ?C T ( ) implying that: xT C ( )x = 0 8 x 2 IR6; 2 IR6

Y m + Y c ; Y d]
"

where:

Ym = Yc= Yd=
"

r _1 0 0 0 r _2 0 0 0 r _3

3. T D( )

2 IR6

0 ?r2 v3 0 r1 v3 0 0 ?r1 v2 r2 v1 0

"

jr1 jv1 0
0 0

B. NPS AUV II MODEL PARAMETERS The AUV model used in the case studies is based on the Naval Postgraduate School (NPS) AUV II, Healey and Lienard (1993). The horizontal motion (surge, sway and yaw) of the NPS AUV II is described by:

0 0 0 v2 0 v 3 0 0 v 3 0 v2

IFAC Workshop on Control Applications in Marine Systems (CAMS'95)