Proceeding of the 2004 American Control Conference Boston, Massachusetts June 30 - July 2, 2004

WeP18.1

On Integral Control in Backstepping: Analysis of Different Techniques

Roger Skjetne1,2 for Ships and Ocean Structures Norwegian University of Science and Technology NO-7491 Trondheim, Norway E-mail: [email protected]
Abstract Including integral action in a nonlinear backstepping design is the topic of this paper. Two methods for adding integral feedback are proposed and analyzed. These are compared to the more traditional methods: 1) adaptive backstepping, and 2) plant augmentation that adds an extra relative degree and thus gives one extra step of backstepping. A test plant is used to compare the different control laws. Based on the theoretical analysis and the simulations, some interesting conclusions are made for each integral control strategy.
1 Centre

Thor I. Fossen1,2 of Engineering Cybernetics Norwegian University of Science and Technology NO-7491 Trondheim, Norway E-mail: [email protected]
tracking error. We therefore consider a class of nonlinear mechanical systems given by the vector relative degree 2 plant x 1 = f1 (x1 ) + G1 (x1 )x2 x 2 = f2 (x1 , x2 ) + G2 (x1 , x2 )u + b (1a) (1b)
2 Department

I. I NTRODUCTION Integral control is one of the principle components in feedback control for industrial use. In theory it has the capacity to remove constant steady-state offsets in a closedloop regulation system. In practice it is a robustifying part of the feedback controller, alleviating problems with unmodeled dynamics, parameter deviations, and slowly varying disturbances. Already one of the rst commercially available feedback controllers, the ship autopilot MetalMike designed by Elmer Sperry in 1922, included integral action to remove offsets in the heading error; see [1], [2]. In fact, through the theoretical analysis of [3] this controller paved the way for what later is known as the PID controller. In modern state-of-the-art industrial control designs for ships, integral feedback is a requirement. See for instance [4] where an LQG feedback controller with integral action is designed and tested in full scale for dynamic positioning of a supply vessel. More recently, nonlinear control designs with integral action included, have been implemented for ship control. The paper [5] is an example where integral action by adaptation is used to counteract slowly varying environmental forces. For linear SISO systems a single integrator can be augmented to the transfer function to compensate a single bias term. For MIMO systems with m inputs and m outputs, this generalizes to m integrators that can be augmented to compensate m bias terms. In nonlinear control systems, feedback laws are often designed by the aid of a control Lyapunov function (CLF). For mechanical systems, this often results in a nonlinear PD-type control law; see for instance [6]. However, awareness to the importance of integral effect is clearly made in [7], [8], [9], [5], [10]. The purpose of this paper is to present and analyze different methods for integral action in backstepping to robustly deal with a bias and to ensure zero steady-state
0-7803-8335-4/04/$17.00 2004 AACC

where x1 , x2 are the states, u is the control input, the functions f1 , f2 , G1 , G2 are smooth, and b is an unknown constant bias. The plant is called undisturbed if b = 0. It is assumed that G1 (x1 ) and G2 (x1 , x2 ) are nonsingular for all x1 , x2 , and all variables are of the same dimension, that is, x1 , x2 , u, b Rm . The plant is a strict feedback-form system representing a class of fully actuated m-degreesof-freedom (m-DOF) mechanical systems. Examples are 3DOF ocean surface vessels, 6-DOF autonomous underwater vehicles, or m-DOF robotic manipulators. For a mechanical system, (1a) typically represents the perfectly known kinematic equation, and (1b) represents the more uncertain dynamic equation. The bias b may represent constant (or slowly varying) environmental forces or perhaps steadystate offsets necessary to maintain the desired equilibrium. The most common way to include integral action in backstepping is to use parameter adaptation [11]. Another method is to augment the plant dynamics with the integral = x1 xd (t). Together with (1), the resulting state system is still in strict feedback form; however, the vector relative degree is increased to 3 and three steps of backstepping is therefore necessary. Based on the complexity of the nonlinear functions f1 and G1 this may involve cumbersome differentiation of two virtual controls resulting in a complex nonlinear control law. The next section will present a design method which avoids a three step design. This results in a negative semi-denite CLF time derivative. Usually, this must be analyzed further by KrasowskiiLaSalles principles or Barbalats Lemma. Recently, a new version of Matrosovs Theorem [12] has been developed in [13]. This is a convenient tool used in this paper to directly guarantee Uniform Global Asymptotic Stability of the closed-loop systems. In fact, a predecessor to this new theorem is [10] where integral action is the main motivation. Notation: In GS, LAS, LES, UGAS, UGES, etc., stands G for Global, L for Local, S for Stable, U for Uniform, A for Asymptotic, and E for Exponential. Total time derivatives of x(t) are denoted x, x , x(3) , . . . , x(n) , while
1899

a superscript denotes partial differentiation: f t (x, y, t) := f 2f n f x2 yn t , f (x, , t) := x2 , and (x, , t) := yn , etc. The Euclidean vector norm is |x| := (x> x)1/2 , while ||x|| denotes the ess sup{|x(t)| : t 0}.The induced norm of a matrix A is denoted ||A||, and col(x, y ) means the column vector of x and y stacked on top of each other. A. Motivational Examples Example 1: A time-varying closed-loop system For the scalar plant let the task be to track the signal xd (t) without steadystate error regardless of an unknown constant bias b. Additionally, the monotone damping term x3 should not be cancelled directly since this may result in unwanted transients in the control law. Under the assumption that b = 0, the control law 1 u= (x xd (t)) + xd (t)3 + x d (t) (3) 2 (1 + x ) x = (1 + x2 )u x3 + b (2)

x1 xd (t) gives proportional action (P), while feedback from x2 gives derivative action (D). We augment the state = x1 xd (t) so that the plant space with the integrator becomes a relative degree 3 strict feedback form system. Three steps of backstepping on the resulting system is the most common way to design a nonlinear PID controller; see for instance [15, Chapter 7.4.5]. A problem with this procedure is that it will attempt to drive the extra integrator state to zero. For b = 0 this will certainly solve the tracking objective. However, we study what happens in the closedloop for b 6= 0. Dene z1 := x1 1 (, x1 , t) and z2 := x2 2 (, x1 , x2 , t) where 1 and 2 are virtual control functions to be specied. Letting 1 = k + xd (t) 1 c1 z1 + x3 d (t) 2 = 1 k (x1 xd (t)) + x g (x1 )

then

renders the equilibrium e = x xd (t) = 0 UGES. This is 2 veried by the Lyapunov function V1 (x, t) := 1 2 (x xd (t)) for which the time derivative satisfy V1 (x xd (t))2 by using the property (x y )(x3 y3 ) 0, x, y. If on the other hand b 6= 0, then the closed-loop equilibrium is shifted with the result that zero tracking error cannot be achieved. The above control law consists of a proportional feedback (P) term (x xd (t)) and a feed-forward (FF) term xd (t)3 + x d (t). To R tensure zero tracking error, the integral equation (t) = 0 (x( ) xd ( ))d is introduced, and (3) is modied with the inclusion of the integral feedback (I) term k where k > 0. This results in a nonlinear PI+FF control law. The closed-loop system = x xd (t) d (t) + b x = k (x xd (t)) x3 xd (t)3 + x has an invariant manifold Mt = {(x, ) : x = xd (t), = 1 k b} on which the tracking error is zero. To analyze stability, 1 2 we dene V2 (x, , t) := V1 (x, t) + 1 2 k ( k b) . This gives 2 (x xd (t))2 0 V

The stable equilibrium zeq = 0 cease to exist for b 6= 0. However, since the unforced closed-loop system is UGES and b enters linearly, the system is input-to-state stable (ISS) with b as input and linear gain dependent on k, c1 , c2 ; see [16], and thus the solutions will stay bounded. If the term x1 (t) = z1 k + xd (t) is time-varying, the steady-state solution is time-varying. If, on the other hand, xd (t) = xref is constant then there exist a new constant equilibrium given by zeq = A(xref )1 eb. In this equilibrium, the tracking error x1 xref = 0 as desired, while the integral state is driven to the non-zero value bg(xref )/(c2 + kc1 c2 + kg(xref )2 ). II. I NTEGRAL C ONTROL Consider the plant (1) and let the control objective be to solve the tracking problem limt [x1 (t) xd (t)] = 0 where xd (t) is a bounded smooth reference signal. This should be achieved in presence of a constant bias b 6= 0. To counteract the effect of the constant unkown bias, we consider using integral action and investigate two methods next, denoted A and B. With the backstepping state transformation z1 := x1 xd (t) and z2 := x2 1 where 1 is a virtual control, Method R t A will add feedback from the integral term (t) = 0 z1 ( )d in the rst step of backstepping. This is perhaps the most intuitive method. However, it will be shown that for a generic plant model this method cannot guarantee convergence of the tracking error, even when the reference signal is constant. To overcome this problem, Method B will instead use feedback from Rt (t) = 0 z2 ( )d in the second step of backstepping. This

is a nonlinear PID control law which renders the equilibrium (, z1 , z2 ) = 0 UGES for b = 0. With z := col(, z1 , z2 ) the closed-loop (time-varying) system becomes z = A(x1 (t))z + eb where e = [0, 0, 1]> and " # k 1 0 c1 g (x1 (t)) . A (x1 (t)) = 1 0 g (x1 (t)) c2

u = g (x1 )z1 c2 z ) [x1 xd (t)] 2 + 2 (, x1 , x2 , t x1 3 t ) x x +2 (, x1 , x2 , t) (1 + x2 2 1 1 + 2 (, x1 , x2 , t)

which is only negative semi-denite. This shows directly that Mt is UGS, but not necessarily UGAS. Since the closed-loop system is time-varying, Krasovskii-LaSalles principles cannot be applied to show convergence to Mt . Instead, one must resort to Barbalats Lemma [14] or some other theorem to prove UGAS. In the next section, application of Matrosovs Theorem [13] will prove UGAS of Mt directly. Example 2: PID control by 3 steps of backstepping Consider the relative degree 2 plant x 1 = g (x1 )x2 x3 1 x 2 = u + b (4)

where the function g (x1 ) is strictly nonzero. We can think of x1 as a position that we want to steer to a desired position xd (t), and x2 as a velocity. Accordingly, feedback from

1900

ensures that is matched with b in the closed-loop and can therefore asymptotically cancel it. Method A, Step 1: For the plant = z1 choose the rst CLF as 1 1 > V1A (, x1 , t) := > KA + z1 z1 2 2 > > 0. The time derivative becomes where KA = KA > > 1A = z1 [KA + f1 + G1 1 x d ] + z1 G1 z2 V and the virtual control are chosen as
> C1 1 d] 1 (, x1 , t) = G 1 [KA C1 z1 f1 + x

d (t) z 1 = f1 (x1 ) + G1 (x1 )z2 + G1 (x1 )1 x

(5) (6) (7)

and the closed-loop system = z1 z 1 = KA C1 z1 + G1 (x1 (t))z2 z 2 = G1 (x1 (t))> z1 C2 z2 + b.

(17)

(8)

which for z2 = 0 is only negative semidenite. For now we allow this and continue the design. Method A, Step 2: Differentiating z2 with respect to time gives where z 2 = f2 (x1 , x2 ) + G2 (x1 , x2 )u 1 +b
x1 t 1 = 1 z1 + 1 [f1 + G1 x2 ] + 1 .

> 0. Recall that good nonlinearities in where C1 = f1 can be exploited at this point when designing 1 , as was the case in Example 1. The above choice yields > > 1A = z1 V C1 z1 + z1 G1 z2 (9)

Proposition 1: The equilibrium (, z1 , z2 ) = 0 of the closed-loop system (17) with b = 0 is UGAS. Proof: To prove stability for the case b = 0 we apply Theorem 1 by [13]. The origin (, z1 , z2 ) = 0 is UGS by (12) and (16). Dene W1 := V2A and W2 := > z1 , and > > > accordingly Y1 := z1 C1 z1 z2 C2 z2 and Y2 := z1 z1 > > > KA C1 z1 + G1 (x1 (t))z2 . Then W1 = Y1 and 2 = Y2 . From the boundedness of xd (t) and continuity W of G1 (x1 ) we get that (z1 , t) := G1 (z1 + xd (t)) and Wi (, z1 , z2 ), i = 1, 2, are bounded for all bounded values of (, z1 , z2 ). Moreover, for 6= 0 then Y1 = 0 Y2 < 0 and Y1 = Y2 = 0 (, z1 , z2 ) = 0. This proves that the origin (, z1 , z2 ) = 0 is UGAS for b = 0.

(10) (11)

With the choice 1 > z2 , V2A (, x1 , x2 , t) := V1A + z2 2 the derivative along the state solutions becomes > > 2A = z1 C1 z1 + z2 b V > > + z2 G1 z1 + f2 + G2 u 1 (12)

(13) (14)

For the case b 6= 0, the closed-loop is a UGAS system forced by the constant input b. Such a constant input may destabilize the system; see for instance [14, p. 177]. Investigating (17) shows that if xd (t) is time-varying, then this system in general cannot settle at a constant equilibrium due to the time-varying term G1 (x1 (t)). As a result, convergence z1 (t) 0 must fail. If, on the other hand, xd (t) = xref =constant then there exist a constant equilibrium for the system - which may or may not be UGAS. Remark 1: In the case when G1 is a constant matrix, then the closed-loop (17) becomes linear and can be written z = Az + eb, where A is Hurwitz and e = [0, 0, 1]> . For b 6= 0, the new equilibrium becomes z0 = A1 eb. In this equilibrium the tracking error is zero. By dening w := z z0 we get w = Aw which shows that z0 is a UGES equilibrium. This conclusion also holds for the 3 step design in Example 2 if g (x1 ) is a constant. Method B, Step 1: For the equation choose the rst CLF as 1 > V1B (x1 , t) := z1 z1 . 2 The time derivative becomes > > 1B = z1 [f1 + G1 1 x d ] + z1 G1 z2 V and the virtual control are chosen as
> C1 1 d] 1 (x1 , t) = G 1 [C1 z1 f1 + x

and the control u is chosen as 1 G> 1 u = G 1 z1 C2 z2 f2 + 2

z 1 = f1 (x1 ) + G1 (x1 )z2 + G1 (x1 )1 x d (t),

(18) (19)

where

> where C2 = C2 > 0. To see that this is in fact a nonlinear PID+FF control law, we state it in the original coordinates as u = KI (x1 , x 2 ) KP (x1 , x2 ) [x1 xd (t)] d f1 (x1 )) KD (x2 ) x2 G1 (x1 )1 (x (15) +FF (, x1 , x2 , t) 1 1 KI (x1 , x2 ) := G 2 C2 G1 KA 1 > 1 1 KP (x1 , x2 ) := G 2 G1 + G2 C2 G1 C1 1 1 + G 2 G1 KA 1 KD (x2 ) := G 2 C2 1 1 x1 FF (, x1 , x2 , t) := G 2 f2 + G2 1 [f1 + G1 x2 ] 1 t + G 2 1 .

(20)

> 0. Good nonlinearities in f1 can be where C1 = exploited at this point when designing 1 . The above choice yields > > 1B = z1 C1 z1 + z1 G1 z2 (21) V which for z2 = 0 is negative denite. Method B, Step 2: Introducing the integral term and differentiating z2 with respect to time gives = z2 (22) 1 +b (23) z 2 = f2 (x1 , x2 ) + G2 (x1 , x2 )u

This control law yields > > > 2A = z1 C1 z1 z2 C2 z2 + z2 b V

(16)

1901

where

t 1 1 = x 1 [f1 + G1 x2 ] + 1 .

(24)

With the choice 1 1 > z2 V2B (, x1 , x2 , t) := V1B + > KB + z2 2 2 > > 0, and the control where KB = KB 1 > u = G2 G1 z1 KB C2 z2 f2 + 1 ,
> C2

(25)

(26) (27)

the derivative along the state solutions becomes > > > 2B = z1 V C1 z1 z2 C2 z2 + z2 b

The control laws are called Method A, Method B, and 3 Step according to the designs in Section II and in Example 2. In addition, an ISS-backstepping control law (see [17] and references therein) is implemented for comparison. Since the adaptive backstepping control law yields the exact same responses as for Method B, this has been left out. However, advantages and drawbacks of the adaptive backstepping control law relative to the other control laws tested here will be discussed below. The objective is for x1 to asymptotically track a desired reference signal xd (t), and the resulting tracking control laws are shown in Table I.
TABLE I I NTEGRAL AND ISS C ONTROL L AWS for the benchmark testplant

Proposition 2: The equilibrium (, z1 , z2 ) = 0 of the closed-loop system (28) with b = 0 is UGAS. Proof: Recall Matrosovs Theorem as stated in [13, Theorem 1]. For b = 0, the origin (, z1 , z2 ) = 0 is UGS by (25) and (16). Dene W1 := V2B and W2 := > > C1 z1 z2 C2 z2 and > z2 , and accordingly, Y1 := z1 > > > > > Y2 := z2 z2 G1 (x1 (t)) z1 KB C2 z2 . Then 2 = Y2 . From the boundedness of xd (t) and 1 = Y1 and W W continuity of G1 (x1 ) we get that (z1 , t) := G1 (z1 + xd (t)) and Wi (, z1 , z2 ), i = 1, 2, are bounded for bounded (, z1 , z2 ). Moreover, for 6= 0 then Y1 = 0 Y2 < 0 and Y1 = Y2 = 0 (, z1 , z2 ) = 0. This proves that the origin (, z1 , z2 ) = 0 is UGAS.

> 0. By writing the control law in the where C2 = original coordinates, one can again verify that it indeed is a nonlinear PID+FF control law. The closed-loop system becomes z 1 = C1 z1 + G1 (x1 (t))z2 = z2 (28) z 2 = G1 (x1 (t))> z1 KB C2 z2 + b.

The advantage of this method appears for the case when b 6= 0. This gives the new (constant) equilibrium 1 1 b, 0, 0). Letting = KB b yields (, z1 , z2 ) = (KB z 1 = C1 z1 + G1 (x1 (t))z2 = z2 z 2 = G1 (x1 (t))> z1 KB C2 z2 (29)

Method A: = z1 z1 := x1 1 xd , z2 := x2 3 k c 1 = 1 z + x ( t ) +x d (t) 1 1 d g x1 t gx2 x3 u = gz1 c2 z2 + 1 + 1 1 z1 + 1 Method B: (equivalent to adaptive backstepping) = z2 z1 := x1 xd , z2 := x2 1 3 1 = 1 d (t)] g [c1 z1 + xd (t) + x t 1 u = k gz1 c2 z2 + x gx2 x3 1 + 1 1 3 Step: = x1 xd (t) z1 := x1 1 , z2 := x2 2 1 = k + xd (t) 3 2 = 1 d (t) + 1 (x1 xd (t))] g [ c1 z1 + x1 + x u = gz1 c2 z2 + 3 2 (x1 t xd (t)) 1 gx x +x 2 1 + 2 2 ISS backstepping: z1 := x1 x2 1 xd , z2 := 3 c 1 = 1 z + x ( t ) +x d (t) 1 1 d g t 1 gx2 x3 u = gz1 c2 z2 kz2 + x 1 + 1 1 A. Simulation Results In the following simulations, the feedback gains was set to k = 0.25, c1 = 1.5, and c2 = 0.5. The closed-loops were tested for a constant reference xd = xref = 1 and also for a sinusoidal reference xd (t) = 1 + 0.5 sin(0.2t) fed through a reference lter to produce the necessary derivatives. The rst test is a step response (xref = 1) using g(x1 ) = 1 + x2 1 and b = 0. The responses for x1 (t) are shown in Figure 1. It is seen that the ISS-backstepping design has a superior response in this case. The other control loops experience a transient in the integral term that requires time to converge. Notice also the indication that x1 (t) for Method A converges very slowly to the reference. In the second test, the bias is turned on, b = 2, and it is clearly seen in Figure 2 that this is detrimental to the Method A and 3 Step closed loops. While the 3 Step rapidly enters a bounded oscillation, the Method B are trying to converge before it also suddenly starts oscillating. The ISS-backstepping design obtains a steady-state
1902

which is the same as (28) for b = 0. By Proposition 2 this gives the following result:
1 Theorem 3: The equilibrium (, z1 , z2 ) = (KB b, 0, 0) of the closed-loop system (28) is UGAS for any constant b.

In other words, this method guarantees zero tracking error for the generic plant (1), even when the reference is timevarying. Remark 2: By designing an adaptive control law using adaptive backstepping [11], one obtains the same controller as in Method B. Indeed, let b be the bias estimate. Then the adaptive closed loop becomes equal to (28) by setting b = KB . III. A C OMPARISON B ETWEEN D IFFERENT D ESIGNS In this section a comparative study is performed on different integral control strategies for the plant x 1 = g (x1 )x2 x3 1 x 2 = u + b. (30)

Time responses for x (t) 1

Time responses for x (t) 1 2.5

1.2

0.8

1.5
0.6

0.4

0.2

Desired Method A Method B 3 Step ISS 0 5 Seconds 10 15

0.5 Desired Method A Method B 3 Step ISS 0 0 10 20 30 40 50 Seconds 60 70 80 90 100

Fig. 1. Step responses (xref = 1) for the closed-loop systems with g (x1 ) = 1 + x2 1 and b = 0.

Fig. 3. Responses to xd (t) = 1 + 0.5 sin(0.2t) with g (x1 ) = 1 + x2 1 and b = 2.

Time responses for x1(t) 2

offset, as expected, but this is preferable to the oscillating responses. Only Method B satises the control objective with success.
Time responses for x1(t) 2.5 2 1.5 Desired Method A Method B ISS

1.8

1.6

1.4

1.2

0.8

1
0.6

0.5
0.4

Desired Method A Method B 3 Step ISS

50

100

150

0.2

2.5 2 1.5 1 0.5 0 Desired 3 Step

10

20

30 Seconds

40

50

60

70

Fig. 4. b = 2.
0 50 Seconds 100 150

Responses to xd (t) = 1 + 0.5 sin(0.2t) with g (x1 ) = 1 and

Fig. 2. Step responses (xref = 1) for g (x1 ) = 1 + x2 1 and b = 2. Upper plot: Method A, B, and ISS. Lower plot: 3 Sted Design.

When the reference is a sinusoid, we observe that the Method B response still converges nicely as shown in Figure 3. The other closed-loops clearly fails. On the other hand, when the function g(x1 ) is constant, then Figure 4 veries the stability and convergence properties of all designed closed loops; see Remark 1. Remark 3: Simulations have shown that the incapacity of Method A and 3 Step to counteract a bias, even when the reference is constant, is gain dependent. Adjusting the positive feedback gains k, c1 , and c2 (usually increasing them) will eventually give stability and convergence. A quantitative analysis of this is future work. B. Discussion The only integral control design presented here that for a generic plant model (1) guarantees UGAS with convergence

of the tracking error in presence of a constant bias b is Method B, which is equivalent to the feedback law obtained by adaptive backstepping [11]. In the plant considered, only one bias is included with constant unity regressor, that is, b = 1b. Adaptive backstepping is a much stronger design methodology since it can handle unknown biases in all state equations, and also more general regressors, for instance, (x)> b. However, adding pure integral action in an already available nonlinear control law is often sought for since it is believed to be a robustifying term that can handle disturbances and dynamics that have not been included in the original design model. Doing so has shown to be applicable by the help of Matrosovs Theorem. However, as illustrated in the above simulations, such designs must be performed with care. In special cases (for instance when G1 (x1 ) is constant) the integral designs will robustly deal with constant disturbances. However, for other plant models (for instance when G1 (x1 ) = 1 + x2 1 ) then UGAS of the undisturbed loop does not imply stability and convergence in presence of constant disturbances. In fact, certain feedback gains that guarantee UGAS in the undisturbed case may give large oscillations for nonzero but
1903

small bias terms. This is the weakness of control laws that only guarantee a UGAS undisturbed closed loop: UGAS do not imply ISS. The properties of each integral control method are summarized below: Method A: In general UGAS closed-loop for b = 0, but no robustness is guaranteed for b 6= 0. When G1 (x1 ) is constant it gives a UGES closedloop for any constant b. When xd is constant it seems to provide stability and convergence for higher feedback gains (no proof). No convergence in general when xd = xd (t). Can be mixed with other designs, for instance, adaptive control to give adaptive PID control or ISS backstepping to give ISS PID control. Method B: Guarantees UGAS and convergence of tracking error in presence of any constant b in all considered cases. Adaptive backstepping by tuning function design gives the same controller. Specialized for the plant (1), that is, matched uncertainty with constant regressor. 3 Step: A 3 step backstepping design is more cumbersome. In general UGES closed-loop for b = 0, and robust with respect to b since the closed-loop is ISS. When G1 (x1 ) is constant it gives a UGES closedloop for any constant b. When xd is constant it seems to provide stability and convergence for higher feedback gains (no proof). No convergence in general when xd = xd (t). Adaptive-backstepping: The most general method since it guarantees stability and tracking error convergence for multiple biases and regressor/bias structures. Involves in general a rather complex nonlinear design, and it is specialized to deal with unknown constant model parameters. It has otherwise the same conclusions as for Method B. ISS-backstepping: Guarantees bounded solutions with linear gain from b to x1 xd (t). Increasing the disturbance damping may increase control effort. Superior performance in undisturbed case, but results in steady-state error in general. IV. C ONCLUSION This paper has elaborated on several designs for integral action in a nonlinear backstepping design. A class of vector relative degree 2 systems with a matched bias

was considered. For such systems, integral action can be included in the design either by adaptive backstepping or by a three step design on an augmented plant model. This paper has shown that integral action can be included at any convenient location in the closed-loop by the help of a control Lyapunov function for which the time derivative only becomes negative semidenite. Matrosovs Theorem is then applied to show UGAS of the overall closed-loop. What variable to take integral feedback from (here z1 or z2 ) must be decided in case by case. In this paper, two such feedbacks (Methods A and B) were proposed and analyzed. Though derived differently, Method B was shown to be equivalent to adaptive backstepping for the considered plant. While Method B solved the control objective with success in the simulation case, Method A experienced undesirable behavior. This illustrated that care must be taken when designing integral feedback since in worst case it may hurt more than it helps. R EFERENCES
[1] S. Bennet, A History of Control Engineering 1800-1930. Peter Peregrinus Ltd., 1979. [2] Sperry Marine, A short history of sperry marine. Internet, accessed Aug. 2003. <http://www.litton-marine.com/companyinformation_sperry-history.asp>. [3] N. Minorsky, Directional stability of automatically steered bodies, J. Am. Soc. Naval Eng., vol. 34, p. 284, 1922. [4] A. J. Srensen, S. I. Sagatun, and T. I. Fossen, Design of a dynamic positioning system using model-based control, Contr. Eng. Practice, vol. 4, no. 3, pp. 359368, 1996. [5] T. I. Fossen and J. P. Strand, Nonlinear passive weather optimal positioning control (wopc) system for ships and rigs: experimental results, Automatica, vol. 37, no. 5, pp. 701715, 2001. [6] B. Paden and R. Panja, Globally asymptotically stable PD+ controller for robot manipulators, Int. J. of Control, vol. 47, no. 6, pp. 16971712, 1988. [7] I. Kanellakopoulos and P. T. Krein, Integral-action nonlinear control of induction motors, in Proc. 12th IFAC World Congress Automatic Control, (Sydney, Australia), pp. 251254, IFAC, July 1993. [8] Y. Tan, J. Chang, H. Tan, and J. Hu, Integral backstepping control and experimental implementation for motion system, in Proc. IEEE Int. Conf. Contr. Appl., (Anchorage, Alaska, USA), pp. 367372, IEEE, Sept. 2000. [9] R. Ortega, A. Lora, P. J. Nicklasson, and H. Sira-Ramrez, Passivitybased Control of Euler-Lagrange Systems. London: Springer-Verlag, 1998. [10] T. I. Fossen, A. Lora, and A. R. Teel, A theorem for UGAS and ULES of (passive) nonautonomous systems: robust control of mechanical systems and ships, Int. J. Robust Nonlinear Contr., vol. 11, pp. 95108, 2001. [11] M. Krsti c, I. Kanellakopoulos, and P. Kokotovi c, Nonlinear and Adaptive Control Design. New York: John Wiley & Sons Ltd, 1995. [12] V. M. Matrosov, On the stability of motion, J. Appl. Math. Mech., vol. 26, no. 5, pp. 13371353, 1962. [13] A. Lora, E. Panteley, D. Popovi c, and A. R. Teel, An extension of Matrosovs theorem with application to stabilization of nonholonomic control systems, in Proc. 41st IEEE Conf. Decision & Control, (Las Vegas, NV, USA), pp. 15281533, IEEE, Dec. 2002. [14] H. K. Khalil, Nonlinear Systems. New Jersey: Prentice-Hall, Inc, 3 ed., 2002. [15] T. I. Fossen, Marine Control Systems: Guidance, Navigation, and Control of Ships, Rigs and Underwater Vehicles. Trondheim, Norway: Marine Cybernetics, 2002. [16] E. D. Sontag and Y. Wang, On characterizations of the input-to-state stability property, Systems Contr. Lett., vol. 24, no. 5, pp. 351359, 1995. [17] R. Freeman and P. Kokotovi c, Robust nonlinear control design. Systems & Control: Foundations & Applications, Boston, MA: Birkhuser Boston Inc., 1996. State-space and Lyapunov techniques.

1904