Proceedings of the 27th Conference on Decision and Control Austin, Texas December 1988

TP10 = 2:30

ADAPTIVE MOTION CONTROL OF RIGID ROBOTS: A TUTORIAL

Romeo Ortegal Mark W. Spong2 Coordinated Science Laboratory University of Illinois at Urbana-Champaign 1101 W. Springfield Avenue Urbana, Illinois 61801

ABSTRACT In this paper, we give a tutorial account of several of the most recent adaptive control results for rigid robot manipulators. Our intent is to lend some perspective to the growing list of adaptive control results for manipulators by providing a unified framework for comparison of those adaptive control algorithms which have been shown to be globally convergent. In most cases we are able to simplify the derivations and proofs of these results as well. s 1. PRELIMINARIES

cial case of the notion of feedback linearization of nonlinear systems. The control objective, which is achievable in the case when the parameters are known, is to obtain a closed loop system which is linear and decoupled. The update law in Craig, et. al. (1986), Spong and Ortega (1988), and Amestegui, et. al. (1987) is driven by a tracking error while Middleton and Goodwin, (1988) use a prediction error. The second class of result that we discuss, Slotine and Li (1986), Sadegh and Horowitz (1987), is conceptually different from the adaptive inverse dynamics in that the control objective is not feedback linearization but only preservation of the passivity properties of the rigid robot in the closed loop. To obtain this objective these results effectively exploit the Hamiltonian structure of rigid robot dynamics. The final result that we discuss, Kelly and Carelli (1988), is intermediate between the other two classes, in the sense that an inverse dynamics controller is used together with the addition of a term that allows preservation of the passivity of the closed loop system. The choice of this additional feedback is again dictated by the Hamiltonian structure of the dynamic equations of motion. The last three papers mentioned above all use a tracking error driven adaptation law. The present paper is organized as follows. In Section 2 we discuss the Lagrangian formulation of rigid robot dynamics, and describe some of the key structural properties of robot dynamics used in adaptive control. This is done to make the paper selfcontained and to set the notation for the remainder of the paper. Section 3 discusses the adaptive inverse dynamics results of Craig, et. al. (1986), Spong and Ortega (1988), Amestegui, et. aI (1987), and Middleton and Goodwin (1988), while Section 4 discusses the results of Slotine and Li (1986), Sadegh and Horowitz (1987), and Kelly and Carelli (1988). In each case we first discuss the algorithm in the known parameter case. The known parameter case provides a best case upper bound of the performance obtainable. Indeed, there is little hope to derive stable adaptive control laws if the known parameter case is unstable or otherwise poorly behaved. For the most part the proofs of stability of the adaptive control laws are taken from the appropriate references, but, in some cases we have been able to simplify and/or streamline the proofs. Finally, in Section 5, we discuss some extensions and open problems. 1.2 Notation and Terminology In what follows we will use the following standard notation and terminology (Desoer and Vidyasagar (1975)):

1.1 Introduction
The problem of designing adaptive control laws for rigid robot manipulators that ensure asymptotic trajectory tracking with boundedness of all internal signals has enticed researchers for many years. Controllers that achieve this objective for all desired trajectories and all initial conditions are said to be globally convergent. It is only recently that adaptive control results have appeared that provide rigorous proofs of global convergence for rigid robots. In this paper we give a tutorial account of some of these results. It is not our intent to survey the bulk of literature in adaptive control of robots. In our opinion, once the existence of globally convergent adaptive control laws for rigid robots has been rigorously established, it is difficult to justify other adaptive schemes based on approximate models, linearizition along nominal trajectories, assumption of slow parameter variation, etc., that do not satisfy the criterion of global convergence.
Ip the control literature there is no universal agreement as to what does and does not constitute an adaptive control algorithm. We have chosen to discuss in this paper only those results that explicitly incorporate parameter estimation in the control law. We have not included adaptive control results which have their basis in the theory of variable structure systems, such as the work of Balestrino, et. al. (1983) and Singh (1985), and the references therein. See also Hsia (1986) for further references.

In this paper we will discuss seven results that satisfy the criterion of global convergence. Adaptive control laws may be classified on the basis of their control objective and the signal that drives the parameter update law. The control objective determines the underlying controller structure whose parameters are to be updated on line. The update law, in turn, may be driven by a signal that measures either the error between the estimated parameters and the true parameters (i.e. prediction error) or the error between the desired output and the actual output (i.e. tracking error). The first four results that we discuss, Craig, et. al. (1986), Spong and Ortega (1988), Amestegui, et. al. (1987), and Middleton and Goodwin (1988), are adaptive inverse dynamics. The term inverse dynamics (also called computed torque) is a spe-

R will denote the set of nonnegative real numbers, and R will denote the usual n-dimensional vector space over R endowed with the Euclidean norm
+

On leave from the National University of Mexico, UNAM-DEPFI, P. 0. Box 70-256, 0 4 . 5 1 0 ,Mexico, D.F. Research partially supported by the National Science Foundation under Grant No. 8516091

88CH2531-2/88/0000-1575$1.OO

1988 IEEE

1575

we have that

(1.1)
We define the standard Lebesgue spaces L, and L2as and

LL(R
and

{ f

:R

II f I I , <0 0 )

such that f is Lebesgue measurable where the LL-norm, 11 f I loo, is defined by

+

+R"

Ilfll,
L;(R
and
+

ess SUP Ilf(t)ll,

(1.2)

@ , 4

II f II < 0 0 )

{ f :R

+ +R" such that f is Lebesgue measurable where the Li-norm, I I f I I2, is defined by
CO

Also

1 ad,, aL = c q q
(1.3)
2 ~ , j

- -*

dP

Ilf112 =

Sllf(t)l12dt.
0

aqk

'qk

Thus the Euler-Lagrange equations can be written

Definition 1.1: A mapping x +.y is said to be passive if and only if

T

<x

I y>T := JxTydt 2 -p
0

k
(1.4)

1, *

,n.

(2.7)

for some p > 0, for all T . We will use the-standard abuse of notation that, if H ( s ) is a transfer function matrix in the Laplace variable s of a (Laplace transformable) function h(t) and r(t) is a function of time, then H ( s ) r will stand for (h c r)(t), where c denotes the convolntion product. With this notation then we can state
\

By interchanging the order of summation in the second term above and taking advantage of symmetry of the inertia matrix, we can show that

Lemma 1.2 (Desoer and Vidyasagar, (1975)) Let

e= ,

H(s)r

(1.5)

where H ( s ) is an nXnz strictly proper, exponentially stable transfer function. Then r g L r implies that e e L i n L k , k L ; , e is continuous, and e -+ 0 as t + 00. If, in addition, r + 0 as -+ CO, then k -+ 0.
2. DYNAMICS OF RIGID ROBOTS

The coefficients

2.1 Euler-Lagrange Equations A standard method for deriving the dynamic equations of mechanical systems is via the so-called Euler-Lagrange equations (see Spong and Vidyasagar (1988) for details)

in (2.9) are known as Christoffel symbols(of the first kind). If we set

(p
=-

ap aqk

(2.11)

then we can write the Euler-Lagrange equations (2.7) as

where, q = (ql, , , . ,q,I) 1s a set of generalized coordinates for the system, L , the Lagrangian, is the difference, K - P , between the kinetic energy K and the potential energy P , and T = (T~, . . . ,T " ) is ~ the vector of generalized forces acting on the system. An important special case, which is true of robot manipnlators, arises when the potential energy P = P ( q ) is independent of 4,and when the kinetic energy is a quadratic function of the vector 6 of the form

T.

- E ~ l l ~ :=~ -4~ D(q)(i ~ l ~ l

1,l

1"

(2.2)

where the nxn "inertia matrix" D ( q ) is symmetric and positive definite for each qER'*. The Euler-Lagrange equations for such a system can be derived as follows. Since L
=

In the above equation, there are three types of terms. The first involve the second derivative of the generalized coordinates. The second are qnadratic terms in the first derivatives of q, where the coefficients may depend on q . These are further classified into two types. Terms involving a product of the type q12 are called centrifugal, while those involving a product of the type qlql where i # j are called Coriolis terms. The third type of terms are those involving only q but not its derivatives. Clearly the latter arise from differentiating the potential energy. It is common to write (2.11) in matrix form as (2.13) C ( S > i ) i + g ( q ) = 7, where the k , j-th element of the matrix C ( q , i )is defined as

Nq)i

K -P

'Cdij(q)qiqj -P(q)
2 i,j

(2.3)

1576

11

ckj

CCijk(q)qi
i-1
"

(2.14)
Wki
Wij

is defined according to (2.13).

Elhii=l

Wkj

-a4j

-)qi
a4k

To show that (2.20) holds independent of the definition of C in (2.13), we define the Hamiltonian H using the so-called Legendre Transformation
H
=

a4i

PTi-L(q,i)

(2.21)

2.2 Fundamental Properties

Although the equations of motion (2.13) are complex, nonlinear equations for all but the simplest robots, they have several fundamental properties which can be exploited to facilitate control system design. We state these properties as Property 1 The inertia matrix D ( q ) is symmetric, positive definite, and both D ( q ) and D(q)-' are uniformly bounded as a function of qcR1'. Property 2 There is an independent control input for each degree of freedom. Property 3 All of the constant parameters of interest such as link masses, moments of inertias, etc., appear as coefficients of known functions of the generalized coordinates. By defining each coefficient as a separate parameter, a linear relationship results so that we may write the dynamic equations (2.13) as (2.15) d q ) = Y(q,i,ii)O = 7 where Y ( q , q , q ) is an n x r matrix of known functions, known as the regressor, and 0 is an r-dimensional vector of parameters. See An, et al. (1985) and Khosla and Kanade (1985). Property 4 Define the matrix N ( q , i ) = h(q) - 2C(q,i). Then N ( q , q ) is skew symmetric, i.e., the components njk of N satisfy nik = -nki. Proof: Given the inertia matrix D ( q ) , the kj-th component of h(q)is given by the chain rule as D(q)ii
+

where L is the Lagrangian and p is the generalized momentum defined as p

=

dL -.

(2.22)

84
Using Lagrange's equations it then follows that H is the sum of the kinetic and potential energies, i.e.,

%Cdij(q)qiqi
ij

P(q)

(2.23)

?hiTD(q)i

P(q).

Again using Lagranges equations, it is easy to derive Hamilton's equations for the system,

. q.
pi

dH 8Pi
--

(2.24)
i- ri

C(Cl,i)i

d H
&li

,i

1, . . . ,n.

(2.25)

Using the above equations (2.24)-(2.25), it follows that the total derivative

dH = dt

d H e-qi
aqi

dH e-pi

iT.

(2.26)

8pi

(2.16) Therefore, the k j-th component of N

=

O n the other hand, from (2.23) it follows that dH ap - = i T D q + l/..iTDi + -6 dt aq .T = q r + ?/*iT(D- 2 C ) i

(2.27)

D-2C is given by (2.17)

nki

d k j - 2cki

where the last equality is obtained by substituting the expression for Dq from (2.13). Thus, comparing (2.26) with (2.27) we have that

iT(h -2C)i
i=l

(2.28)

independent of the manner in which C is defined. The property (2.28) is simply a statement that the so-called fictitious forces, defined by C(q,i)q, do no work on the system. A n important consequence of this is the following: Proposition 2.1 The dynamic equations (2.13) of a rigid robot define a passive mapping r -+ 4, i.e.,
T

Since the inertia matrix D ( q ) is symmetric, i.e., dij = d j i it follows from (2.17) by interchanging the indices k and j that
njk
=

-nkj.

(2.18)

<i IT
Remark: The skew symmetry property 4 is a source of some confusion in the robotics literature and some further clarification is therefore in order. Note that we have defined the matrix C using the Christoffel symbols, from which Property 4 follows easily. Given the equations of motion (2.7), other definitions of C are, of course, possible. For example, inspecting (2.7) we might choose the entries of C, as some authors have done, as adki 1 dd.. Cki = E{-- -+ql. (2.19) 1 8% 2 8 % It turns out, no matter how C is defined in (2.13), it is always true that for some /3

> ~ =

SqTrdt 2-P
0

(2.29)

> 0, for all T .

T

Proof: From (2.26) we have

T
=

Si'rddr
0

SdH
0

H ( T ) -H(O)

2 -H(O)

(2.30)

since H ( T ) is nonnegative for all T . It is this property (2.30), or equivalently (2.26), that is referred to in the literature as the passive structure of rigid robots. Example 2.2 Planar Elbow Manipulator Consider the planar manipulator with two revolute joints shown in Figure 1. Let us fix notation as follows: For i = 1,2, qi denotes the joint angle, which also serves as a generalized coordi1577

iT(h -2C)i
However,

0.

(2.20)

- 2C is itself

skew symmetric only in the case that C

nate; m, denotes the mass of link i, I, denotes the length of link i; I,, denotes the distance from the previous joint to the center of mass of link i; and I, denotes the moment of inertia of link i about an axis coming out of the page, passing through the center of mass of link i. The Euler-Lagrange equations for this system can be shown to be (Spong and V;dyasagar-(1988)):

All of the adaptive control results in this paper rely on Property 1 that the inertia matrix is symmetric, positive definite, Property 2 that there is an independent input for each degree of freedom, and on Property 3 that the system is Linear in the unknown parameters. Therefore we will divide them according to whether or not they also rely on the skew symmetry Property 4.

diiiji + d12ij2 + ~ 1 2 1 4 1 4 2+ C2114241+

d2lijlf d 2 2 i 2 + c1124;

.2

C22142

+ $1

= TI

(2.31) (2.32) (2.33) (2.34) (2.35) (2.36)

(2.37)

3. INVERSE DYNAMICS BASED CONTROL

We first investigate control methods based on the so-called method of inverse dynamics, or computed torque. 3.1 Known Parameter Case Inverse dynamics based control schemes do not exploit the skew symmetry Property 4, but relies instead on exact cancellation of all nonlinearities in the system so that, in the ideal case, the closed loop system is linear and decoupled. Inspecting (2.13) we see that if the control r is chosen as
7 = D(s)a + C(q,Ci)jl + d q ) , then, by substituting (3.1) into (2.13) one obtains

+ $2 = 7 2 .

where
d,,
=

nt11,12 + m,(ll + 1,;

+ Z,~,2cosq2) + I , + I2 ,

d,,

= d,, = 17Z2(1,;

+ l , l , , ~ ~ ~+qI ~ , , ) + I2
- ~ ~ z ~ l , l , ~ s := inq h~
=

d,,

= nt21,,2 =

clzl = cZll = c,,,

=

-cl,,

(/?ZlZCl + nt,l,)gcosq, + f?~21cpJs(q, + q2)

$2 =

~722~,zgC0s(q, + q2).

(34 (3.2) (3.3)

In this case the matrix C(q,i) is given as D(q)(q - a ) and Property 1 implies that
=

(2.38)
and D - 2C is

; i = a.

The term a has the interpretation of an outer loop control law with units of acceleration, which can be defined in terms of a given linear dynamic compensator K ( s ) as

(2.39)
Now if we define parameters O,, . . . ,Og as
0,
8,
8,
= m,l,",
= 11Z21: = 17t21:,

q -K(s)e

..d

(3.4)

with e = q - qd, where qd(t) is an n-dimensional vector of desired joint trajectories. Substituting (3.4) into (3.3) leads to the linear error equation

0,

=in2~l~c2

0,

= nt,l,,g

0,
8,

= =

I,
1,

0, = I7Z21,g
8, = 1 7 1 2 1 c g

(2.40)

The simplest choice of K ( s ) in (3.4) is a PD-compensator K(s)

=

K,s

Kp,

(3.6)

we can write (2.31)-(2.32) as

Y(q
9

which leads to the familiar second order error equation

4,$0
=

= 7

(2.41) (2.42)
qzq,q2-sinq2q2

k' + K,d + Kpe

(3.7)

where 0

[O,,

. . . ,OgJT, and Y is given by

Y(q,4,3

I the gain matrices K,, and Kp are chosen as diagonal matrices

4; 4; 4;+ 41 2cosq24;
0 0 q; + q ;

i cosqZq;-2sin

..

.2

with positive diagonal elements then the closed loop system is linear, decoupled, and exponentially stable. Global stability for this scheme is thus obvious. In fact the closed loop damping ratio and natural frequency may be arbitrarily assigned. The freedom allowed by the compensator K ( s ) in (3.4) may be used to shape the error transients or, as shown in Spong and Vidyasagar (1987), to improve the robustness of the inverse dynamics scheme.

cosqzq; + sinqz4f

3.2 Adaptive Case

Note that the choice of parameters in the above representation is not unique and that the dimension of the parameter space may depend on the particular choice of parameters. Also, not all of the parameters in the system may be unknown or it may be desired to estimate oiily a subset of the parameters. In this case we can write (2.41) as where on contains only known parameters and 0 contains those parameters which are to be estimated. All of the control schemes to follow may be modified to account for this case, although, in the interest of simplicity of the exposition, we will consider only the parameterization (2.41). We review below four different versions of adaptive inverse dynamics control, namely, Craig, et. al. (1986), Spong and Ortega (1988), Amestegui et. al. (1987), and Middleton and Goodwin (1988). We break up the discussion into three parts, according to the assumptions and measurements needed for the implementation of the control law. The first result, Craig, et. al. (1986) requires both measurement of the joint acceleration and modification of the adaptation algorithm to insure boundedness of the inverse of the estimated inertia matrix. The second result, Spong and Ortega (1988), removes the requirement of Craig, et. al. (1986) on boundedness of the estimated inertia matrix. The third result, Amestegui, et. al. (1987), also removes the requirement on boundedness of the estimated inertia matrix, but uses a different parameter update law. The final result in this section, MiddIeton and Goodwin (1988) removes the requirement on measurement of

1578

the joint acceleration but still requires boundedness of the inverse of the estimated inertia matrix. 3.2.1 Control Assuming Acceleration Measurement and Boundedness of the Inverse of the Estimated Inertia The adaptive implementation of (3.1)-(3.4) proposed in Craig, et. al. (1986) is obtained by replacing D , C, and g by their estimates, i.e.,
T=

pleted using the implication

x uniformly continuous
implies x
t

,XEL?

(3.21)

0 as f + 00.

3.2.2 Control Assuming Only Measurement of Acceleration The two main drawbacks of the result in the previous section are the need to measure the acceleration q in order to realize the update law (3.15) and the requirement that fi remain uniformly positive definite. We can remove the second drawback above by following the approach of Spong and Ortega (1988) which is the approach typically used in robust nonadaptive control (see, for example, Spong and Vidyasagar (1987)). In this approach we choose an inverse dynamics control law of the form
7 =

d(q)(qd-K,&KPe)

and p have the same functional form as D , We assume that 8, C, g with estimated parameters g1, . . . Thus

e,

+ t ( q , G ) i + g(q).

(3.8)

,ar.

Dq
where

+ ti + g

Y(q,q,q)8

(3.9)

8 is the vector
Dq

of estimated parameters. B(qd-K,b-KPe)

Substituting (3.8) into (2.13) gives

Do(s)(a

6 4 + C o ( s , i > i + go

(3.22)

+ C4 + g

+ g.

(3.10)

Adding and subtracting Dq on the left hand side of (3.10) and using (3.9) we can write

D(k'
where as

K,,i

Kpe)

= =

bq + c4 + g
Y(q,i,ii)J

(3.11)

where D o = D; > 0, CO, go are a priori estimates of D , C, g, respectively, with fixed parameters, a is given by (3.4), and 6a is an additional outer loop control that compensates for the deviations AD, AC, Ag, where A(.) = In the present setup 6a is chosen adaptively. It is important to note that the terms Do, CO,go in (3.22) are not updated on-line and hence the invertibility of D ois not an issue.
- ( a ) .

(r)

:= (:) - (.). Finally the error dynamics may be written

If we now combine (3.22) with (2.13) we have an equation similar to (3.11)

Do(;

k' + K j

+
k

Kpe

D-'Y8:=08.
B08

(3.12)

+ Kv6 + K p e 4 a ) = ADq + AC4 + Ag

=

(3.23)

We may write the system (3.12) in state space as

=

Y(s,4,q)Ao

Ax

(3.13)

where A is the Hurwitz matrix

where the last equality is obtained using Property 3 of linearity in the parameters. Note, in (3.23), that A O = Bo - 6 ' is a fixed vector in R" and not a function of time, since the terms in (3.22) are fixed estimates. Finally we write

k'
In Craig, et. al. (1986) it is assumed that q is measurable and that the update law is modified to insure that B-' is bounded. See Craig (1988) for further details. Under these assumptions we have the following Theorem 3.1: Choose the update law

+ K v i + Kpe = DO'YAO + 6a:=QOAO + 6a.

6a = -e0A8

(3.24)

Choosing the control 6a as (3.25) yields an equation identical to (3.13) with cf, replaced by a0. Note that A8 = 8 and that now a0is not a function of the estimated parameters since D o is fixed. Choosing an update law for A8 according to

i
ATP

-r-'@TBTPX.

(3.15)

where r = I'T > 0 and P is the unique symmetric positive definite solution to the Lyapunov equation

Ai

-r-'@;BTPx

(3.26)

PA

where P satisfies (3.16), and choosing a Lyapunov function candidate

(3.16)

xTPx

A8'rd

xTPx

aTI'#

(3.27) that

for a given symmetric, positive definite Q. Under these conditions then, the solution x of (3.13) satisfies

a proof identical to that of Theorem 3.1 shows x + 0 as t .--t CO, with all signals remaining bounded.

x - t O as t + c c
with all signals remaining bounded. Proof: Choose the Lyapunov function candidate

(3.17)

This approach is similar to that of Amestegui, et. al. (1987), but the update law is that of Craig, et. al. (1986). The actual control scheme contained in Amestegui, et. al. (1987) is of the form
=

v
v

XTPX

(3.18)

Do(s)a + C,(q,i>ir + go(s) + U

(3.28)

The time derivative of V along trajectories of (3.13) is computed to be

=

where Do, CO,go, a are as in (3.23) and U is an additional signal to be designed. Substituting (3.28) into (2.13) gives

-xT~x

2 i j T p T ~ T+ ~X rj]. -xTQx<0.

(3.19) where (3.20)

ADq

AC4

A g

U =

YAO

(3.29)

Using the parameter update law (3.13) this reduces to

:= D,(v - ti).

(3.30)

This shows that xEL,2"nLE and $ELL. From this we conclude from (3.8) that TELL. This in turn implies, using (2.13) and Property 1, that { E L L , and, hence, from (3.13) that k L : . Since EL:, x is uniformly continuous and the proof is com-

The following adaptive compensation signal is proposed in Amestegui, et. al. (1987)
U =

YbO

(3.31)

1579

with the update law

de

PW (8.) l
=

Bi + --Siw.
W

1.

(3.43)

PYTE

(3.32)

With the preceding control structure we can then prove the following Theorem 3.2 Consider the robot dynamics (2.13) in closed loop with the controller (3.28), (3.31) and update law (3.30), (3.32). Then, assuming that a solution of (3.32) is defined for t E [ t o ,CO), we have e , 6 -+ 0 as t with all signals remaining bounded. Proof: First note from (3.30) and (3.31) that we can write
E =
-+ CO

(3.33)

Hence the implementation is possible with knowledge of 8,. The second modification introduced in Middleton and Goodwin (1988) is needed to account for the noncommutativity of the operator algebra, for example, interchange of filtering and multiplication. This modification is based on the following swapping property 1 d W-'(B-*E) = -[-(B-l)]E + B-lW-'(c). (3.44) w dt The controller is then given by
T =

Ba +

+ 2 + { - Y ~ P + -B[--(B-')~I~
W

(3.45)

dt

-YdO

(3.34)

and consider the Lyapunov function

v
Then V satisfies

where the terms inside the braces are the modifications mentioned above. We note that for implementation purposes the last term of (3.45) may be written more simply using the identity (3.46)

LioTrLie.
5 0,

(3.35)

v
e
=

-dBYTYdO

(3.36)

which implies that E ELin!; and &ELL. On the other hand, using (3.4), (3.30) and invertibility of Do yields

The motivation for the modifications in (3.45) become clear by applying W-' to (3.41)
W-lE
=

YP

+ -YfP - r
W

(3.47)

(s21 + ~ , s+

K,)-'D;'E.

(3.37) ~ ' ( f i - l e )=

Noting that D;'E EL;, it follows from Lemma 1.2 that e -+ 0. In fact, since s(s Z + Kvs + ICp)-' is still strictly proper, stable, it immediately follows from Lemma 1.2 that -+ 0, likewise, and the proof is complete.

which, after substitution of (3.45) gives

(21+

~ , , s + K,)e

(3.48)

3.2.3 Inverse Dynamics Without Joint Acceleration Measurement

The approaches in the preceding section remove the difficulty of retaining invertibility of the estimated inertia matrix, but still require measurement of the acceleration vector q . A reasonable way to remove the acceleration from the regressor is to filter (2.15) to obtain
Tj =

where we have used (3.11), the swapping identity (3.44), and (3.46). The proof that e -+ 0 as t -+ CO follows immediately from Lemma 1.2 since EE L; and 6-l is bounded. A n argument similar to that used in Theorem 3.1 can be used to conclude also that

i -+Oast

-+ 00.

4. PASSIVITY BASED CONTROL METHODS

Yjh4Y
W

(3.38)

where
=

W ( * )= -(.)

, w > 0.

(3.39)

s+w

We next investigate control schemes that exploit the skew symmetry property 4. These results, in general, do not lead to a linear system in the closed loop even in the ideal case that all parameters are known exactly. The main motivation for these schemes is that, as will b e shown, they lead in the adaptive case to error equations where the regressor is independent of the joint acceleration. A s before we first investigate the control of (2.13) in the ideal case that all parameters are completely known. 4.1 A General Theorem The following theorem will be used to unify several of the adaptive control results considered as special cases of a more general result. Theorem 4.1 Let t -+ q d ( t )be a iven twice differentiable func$ tion, and define e ( t ) = q(t) - q ( t ) . Consider the differential equation

Since the acceleration terms appear only in conjunction with

D(q), the ql terms are multiplied by known functions of q only.

Consequently Yf contains no acceleration terms. If we then define a prediction (or input matching) error

E:=Y f
E

8 - 7

(3.40)

will clearly satisfy

E =

YjJ.

(3.41)

Thus any standard parameter update law gives E E L i n L L and &L',. The problem now is how to relate the properties of E to properties of the tracking error e. This problem has been solved in Middleton and Goodwin (1988) with the introduction of two clever modifications to the control law. First, to account for the effect of the filtering, the controller parameters 8, are replaced by

D(q)i. + C(q,G)r + K,r

where D , C are as in (2.13), K,
=

\Ir.

(4.1)

KvT> 0, r is given by
(4.2)
+ \Ir

F(s)-'e

where F ( s ) is strictly proper, stable, and the mapping -r passive, i.e.,

T

is

PW(8,)=

w-'8,w

(3.42)

f-rT(t)\Ir(t)dt 2 -p
0

(4.3)

with W the filter defined by (3.39). This same idea was used in Narendra, et. al. (1980) to prove stability of model reference adaptive control of systems with relative degree two. It is easy to see that

for all T and for some ,8 2 0. Then eEL;nL:,, ;EL;, e is continuous and e -+ 0 as t -+ CO. In addition, if \Ir is bounded, then r -+ 0 as t -+ CO and, consequently, 6 + 0.

1580

Proof Consider the function V defined by

I

pulator acceleration, but only on v and a, which depend on the velocity and acceleration of the reference trajectory.

%rTD(q)r

b-SrT(r)9(r)d7,
0

(4.4)

and note from (4.3) that V tories of (4.1) gives

V
=

2 0.

Differentiating V along trajec-

In order to apply Theorem 4.1 we need to define a parameter update law in such a way that the mapping -r + Q in (4.14) is passive. We therefore choose the parameter adaptation law

j=

-r-'yTr

(4.15)

rTD(q)i.

+ +
=

%rT6(q)r - rr(t)9(t)

(4.5) (4.6)

Substituting fbr D(q); from (4.1) gives

for some symmetric, positive definite matrix (4.14),(4.15) we have

r.

Then, from

-rTKvr

%rr(6(q) -2C(q,i))r -rlKvr

rT9
and, hence,
I
1

rTy8

-e re

:T

(4.16)

by Property 4. Therefore rd,; and it follows from Lemma 1.2 that e + 0 and 6 is bounded. The proof that 6 -+ 0 if 9 is bounded is similar to the proof of Theorem 3.1. Remarks: 1) We note in (4.6) that it is important to use the particular choice of C that makes D - 2C skew symmetric. 2) It is also important to note, since F ( s ) is strictly proper, that r = F(s)-'e contains derivatives of e, and hence, derivatives of q. If F ( s ) has relative degree one, then r contains only the first derivative of e and consequently does not depend on the acceleration ti. 3) An alternative proof of Theorem 4.1 is obtained by invoking the passivity theorem (Desoer and Vidysagar (1975)), and using Lemma 1.2. For further details see Kelly, et. al. (1988).

.T

-SrT9dr
0

PBdr
0
I

(4.17)

d %S-(8'r$)dr 0 dr

v2B(tfr8(t) - v28(o)Tre(o)

2 -v28(ofrl(o)
and the mapping -r + 9 is passive. Therefore r E L i , and e -+ 0 by Theorem 4.1. If we take /3 = I/z8(0)Tr8(O) i n (4.4) and note that V EL,, then we conclude that 8 EL', and r E L L . From (4.14) it follows that 9 ELL and the rest follows from Theorem 4.1. Special Cases From the general result that we have proved (Theorem 4.1) it is now straightforward to recover, and hence unify, the algorithms of Slotine and Li (1987), Sadegh and Horowitz (1987), and Kelly and Carelli (1988). It turns out that the choice K(s)
=

4.2 Known Parameter Case Using Theorem 4.1 we can attempt to find a control law 7 for (2.13) that results in an equation of the form (4.1). In the known parameter case we see that the choice
7 =

o ( d a + C(q,i)v

g(q) - K v ( i - V)

(4.7) (4.8) (4.9) (4.10)

law K(s)

1 -A
S

(4.18)

when substituted into (4.1), results in D(q)F where r is defined as

C(q,q)r

Kvr

in (4.10) where A is a diagonal matrix with positive diagonal elements, leads to the algorithm of Slotine and Li (1987). In this case

r
and

q-v.

F(s)

(sZ + A)-'.

(4.19)

If, on the other hand, we choose an outer loop PID control a = +

where a is given as before by (3.4) so that 1 r = (ST + -K(s))e :=F(s)-'e.

S

Kp

Kds

+ S

KI

(4.20)

(4.11)

we obtain the scheme of Sadegh and Horowitz (1987). The control law proposed in Kelly and Carelli (1988) is given by

Now, if the transfer function K(s) is chosen so that F(s)-' is strictly proper, stable, it follows from Theorem (4.2) that e , 6 + 0. If, in addition F ( s ) has relative degree 1, then implementation of the control law 7 from (4.7) requires only joint position and velocity measurements.

(4.21)
W

4.3 Adaptive Version

The adaptive version of the above result proceeds as follows. Given the system dynamics (2.13) we choose the control law
7 =

with a,, ,q, defined by the filter (3.39). Compare this with (4.7). The motivation behind this scheme is that (4.21) can be written as an inverse dynamics control with a compensation term, i.e., r
=

d(q)a

+
g

Qq,q)v da
=

g(q) - ~ , r

(4.12)

d a

g--&

(4.22)

Substituting this into the system (2.13) gives Dq Now, since q D;

=

+
Cr

Cq

i. + a and 6

(4.13) + & + 2 -Kvr. + v we can write (4.13) as

cv
f

with r, as in (3.39). The error equations for this controller are derived replacing (4.22) in (2.13) Di. or equivalently,

Kvr

da

(4.14)
1

+ ki.,
W

Y8

(4.23)

Y(q,q,v,a)B := \k

Note that the regressor function Y does not depend on the mani-

D(-Ff
W

+ Zf) + -cif
W

YB.

(4.24)

1581

Choosing a gradient update law

4.1 allows us to conclude global stability of the overall adaptive (4.25)

system. It is worth mentioning that the choice of the estimator in the schemes of section 4 is only restricted by the passivity property. Therefore adaptation laws with proportional terms (Landau and Horowitz (1988)) and freezing capabilities (Kelly, et. al. (1988)) also provide globally stable schemes. A n issue that remains to be addressed, possibly with a benchmark example, is the comparison of the relative merits of these different control configurations and estimators. Another problem of particular practical interest is the use of discrete-time update laws to reduce the computational burden. Some results in this respect have been reported in Hsu, et. al. (1987).

and noting that D is bounded, we see that (4.24), (4.25) also fit into the framework of Theorem 4.1. Therefore, global convergence of the adaptive scheme follows.

5. EXTENSIONS AND OPEN PROBLEMS

All of the previous schemes insure asymptotic tracking of a desired reference trajectory for all possible initial conditions and with all external signals remaining bounded. A s shown above, the weakest hypotheses under which global asymptotic stability can be established are that the robot be described by (2.13) and satisfy properties 1-4 ( or properties 1-3 if one is willing to use the (filtered) acceleration). There are, however, several practical issues which remain to be investigated. First, the results say very little about the transient performance, only that the signals remain bounded. Second, since asymptotic stability has not been proved to be uniform, small changes in the dynamics may result in loss of stability. Remember the celebrated Rohrs counterexamples (Rohrs, et. al. (1985)). O n the other hand, small unmodeled bounded disturbances may cause unacceptably large (but bounded) deviations from the desired response (bursting phenomena). A s shown in Reed and Ioannou (1987), suitably tailored bounded disturbances can even drive the system unstable. In this section we will discuss some of the issues involved and some of the possible modifications proposed to alleviate these fundamental problems.

5.3 Robustness To address the robustness problem, two avenues of research have been pursued in the adaptive control literature (see e.g., Ortega and Yu (1987)). These are local analysis for slow adaptation, and the development of robustified estimators. In the first approach it has been shown, e.g., Anderson, et. al. (1986) that useful information of the adaptive system can be obtained by considering small adaptation gains which enforces a time scale separation between the dynamics of the controlled plant and the slowly varying controller parameters. In the authors opinion, the specific nature of the robotics problem, with rapidly changing parameters, makes the slow adaptation approach questionable.
A s opposed to the above approaches, which yield local results, there is also a tendency to develop globally robust adaptive controls for black box systems via the introduction of ad hoc fixes to the estimation law, e.g., normalization, dead zones, forgetting, etc. The theory here is basically qualitative and coarse and concentrates on bounded input bounded state stability, without quantifying the operator gain. Such approaches, although important, overlook such effects as extreme sensitivity to tuning parameters and initial conditions and the presence of unpleasant oscillatory solutions (see, e.g., Praly (1988). Nevertheless we believe that such techniques will be useful for the practical solution of some adaptive robotic control problems if the structure of the uncertainty is suitably exploited in the design. In the context of adaptive control of manipulators it is realistic and relevant to consider the robustness with respect to at least five major perturbations: bounded disturbances, flexibility (particularly joint flexibility), actuator dynamics, and friction. (See Sweet and Good (1984).) Robustness here is understood as the preservation of the stability properties and, it is hoped, control performance as well, under small perturbations. In a recent report Reed and Ioannou (1987) introduced a switching integrator leakage ( u- modification) in the estimator to study the problem of bounded disturbances and unmodeled actuator dynamics. The controller structure of Craig, et. al. (1986) and Slotine and Li (1987a) was considered in the paper. The usual dynamics of a permanent magnet DC-motor were included in the system description, i.e.,

5.1 Persistency of Excitation

The sensitivity to disturbances can be overcome by requiring that the regressor signals be persistently exciting so as to guarantee uniform asymptotic stability of the controller parameters or parameter convergence. Some results along these lines have been reported in Craig et. al., (1986), and Slotine and Li (1987b). For the inverse dynamics schemes of Theorem 3.1 it can be shown that parameter convergence is attained if BTP(sI -A)-lB is is persistently exciting. strictly positive real, and Y(qd,id,qd) Persistency of excitation arguments can also be invoked to determine convergence rates that provide some information on the transient performance. 5.2 Other Update Laws It is widely recognized that, at least locally, the convergence rate of least squares estimates is better than gradient update laws. Unfortunately, least squares estimators do not satisfy the passivity property (4.3) required by the schemes of section 4. In order to overcome this problem Slotine and Li (1987) recently introduced a least squares parameter update law that uses both the tracking error and the prediction error, i.e.,

Ir

-F(Yr + yft)
=

(5.1) (5.2)

F1 =

Yryr ; F ( 0 )

F(o)T > 0.

Interestingly enough, it is possible to show that this estimator defines a map H , : ( - E , r) -+ CYf$,Y j ) with the required input/output properties, i.e.

J,nF

Bm4

Nr R,I

= =

KTI
-Ke4 +
U

(5.4)

V,(T) : = < y f 8I
T

- E > ~

+ <ye I -r>T

L,I

(5.3)

+ /2JIIYfIr112dt 2-3
0

where U , and I are the armature voltage and current, respectively, Jm, B , are the rotor inertia and damping, L a , R, are the armature inductance and resistance, N is the gear ratio, and K,, K, are proportionality constants. The basic idea is to treat the inductive motor time constant

where

= 1/21r(o)TF-(o)Ir(o).

(5.4)

L,/R, as a small parasitic assuming that it is O(p), where p < < 1.

Then invoking the robustness of the switching u-modified estimator (Ioannou and Tsakalis (1986)) to weakly observable parasitics

The estimator together with the error equation (4.14) defines a feedback system. Replacing this functional in (4.4) of Theorem

1582

and bounded disturbances, conditions can be given that ensure boundedness of the solutions. Specifically, on can show, after some straightforward calculations, that (2.1), (5.4)can be written as DAF
rl
= f

In this case Theorem 4.1 is still valid and insures global convergence for friction satisfying
T

Svfdt
0

5 0.

C,b

g
f

(5.5)
(5.6)

ILHl(S)L

PH,(s)F

where H1, H , are strictly proper stable transfer functions such that

Other results on adaptive friction compensation have been reported in Canudas, et. al. (1986).
REFERENCES

and D A , CA include the effective motor inertia and motor damping together with the rigid body equations of the robot. The following result can be established for the perturbed system (5.5)-(5.6) with either one of the controllers mentioned above and the parameters updated with a switching mnodified estimator:
, soluTheorem 5.1 There exists p* > 0 such that for p ~ [ O , p * ]all tions with initial conditions inside a suitable neighborhood of the origin are bounded.

Amestegui, M., Ortega, R., and Ibarra, J.M. (1987), Adaptive Linearizing-Decoupling Robot Control: A Comparative Study of Different Parametrizations, Proc. 5th Yale Workshop on Applications of Adaptive Systems Theory, New Haven, Conn. An, C.H., Atkeson, C.G., and Hollerbach, J.M., (1985), Estimation of Inertial Parameters of Rigid Body Links of Manipulators, Proc. IEEE Conf. Decision and Control, Ft. Lauderdale. Anderson, B. D. O., et al., (1986), Stabiliry of Adaptive System: Passivity and Averaging Analysis, MIT Press. Balestrino, A., De Maria, G., and Sciavicco, L., (1983), An Adaptive Model Following Control for Robotic Manipulators, J. Dyn. Sys., Meas., and Cont., Vol. 105, 143-151. Canudas, C., Astrom, K.J., and Braun, K., (1986), Adaptive Friction Compensation in DC Motor Drives, Proc. IEEE Robotics and AutoSan Francisco. mation Conf.,

Global asymptotic tracking is recovered in the limit as 1-1 --+ 0. A n upperbound on the norm of the robot parameters is required for the estimator implementation. The effect of joint flexibility on existing adaptive controllers is an interesting and highly non-trivial problem. In this case an nlink robot has 2n degrees of freedom, represented by the motor shaft angles and the link angles, which are coupled through the joint flexibility. This system can be modeled as (see Spong (1987)) D(91)il
+

C(qlbl)41

d q , ) + K(q, - 9 2 )

O(5.8)

J F z + B i l - K ( q , - 92)

Craig, J., J.,(1988), Adaptive Control of Mechanical Manipulators, ~ ( 5 . 9 ) Addison-Wesley Publishing Co., Reading, MA. Craig, J.J., Hsu, P1, and Sastry, S.,(1986), Adaptive Control of Mechanical Manipulators, IEEE Int. Coizf. Robotics and Automation, San Francisco, CA, March. Desoer, C., and Vidyasagar, M. (1975), Feedback Systems: InputOutput Properties, Academic Press, New York. Hsia, T. C., Adaptive Control of Robot Manipulators: A Review, (1986), IEEE Int. Conf. Robotics and Automation, San Francisco.
Hsu, P., Bodson, M., Sastry, S., and Paden, B. (1987), Adaptive Identification and Control for Manipulators without Using Joint Accelerations, Proc. IEEE Conf. on Robotics and Automation, Raleigh, NC., 1210-1215.

where q1 and qz represent the link and motor angles, respectively, J represents the actuator inertias and K represents the joint stiffness. In this case there is no longer an independent control input for each degree of freedom and the mapping T + bl is no longer passive. Thus, both Properties 2 and 4 are lost. As a result none of the schemes in this paper can be used to control the flexible joint system (5.8)-(5.9) independent of the joint stiffness. In the case that the joint stiffness is large (i.e., 0(1/?)) Spong (1988) has shown that any of the schemes of this paper may be used to control (5.8)-(5.9) provided a high gain (i.e., 0(1/6 )) damping term is added to the control law to damp out the elastic oscillations due to the joint flexibility. In other words, with a control law of the form

(5.10)
where rr is one of the rigid control laws treated in this paper and K, is 0(1/~), the closed loop system exhibits a two-time scale behavior. The elastic oscillations at the joints decay in a fast time scale as a result of the high gain damping term after which the response of the system is nearly the same as that of a rigid robot with the rigid control law r r .

Ioannou, P. and Tsakalis, K. (1986), A Robust Direct Adaptive Controller, IEEE Trans. Automatic Control, AC-31, No. 11, 1033-1043. Kelly, R., and Carelli, R., (1988), Input-Output Analysis of an Adaptive Computed Torque plus Compensation Control for Manipulators, submitted to 27th IEEE Conf.011 Decision and Control, Austin, Texas. Kelly, R., Ortega, R., and Carelli, R., (1988), Adaptive Motion Control Design of Robot Manipulators: An Input-output Approach, Proc. IEEE Conf. on Robotics and Automation, Philadelphia. Khosla, P., and Kanade, T., (1985), Parameter Identification of Robot Dynamics, IEEE Conf. on Decision and Control, Ft. Lauderdale. Landau, I. and Horowitz, R. , (1988), Synthesis of Adaptive Controllers for Robot Manipulators Using a Passive Feedback Systems Approach, Proc. IEEE Conf on Robotics and Automation, Philadel-

Friction has a significant effect on the performance of many robot manipulators, particularly those with gear reduction, mainly because the functional dependence of the friction on the joint variables is difficult to model. In the case of viscous friction, i.e., affine dependence on q, both the inverse dynamics and passivity based adaptive schemes can easily be modified to account for friction. A less restrictive assumption on the nature of the friction force f is dissipativity, Le,
T

1583

phia. Middleton, R.H., and Goodwin, G.C. (1988), Adaptive Computed Torque Control for Rigid Link Manipulators, Sysienzs arzd Control Letiers, Vol. 10, 9-16. Narendra, K.S., Lin, Y., and Valavani, L. (1980), Stable Adaptive Control Design: Part 11, Proof of Stability, IEEE Trans. Autornatic Control, AC-25, 440-448. Ortega, R. and T. Yu, (1987), Robustness of direct adaptive controllers: a survey, 10th IFAC World Corzf., Munich, FRG. Praly, L., (1988), Oscillatory Behavior and Fixes in Adaptive Linear Control, A Worked Example, CA1 Ecole de Mines Rapp. Iiii. Reed, J.S., and Ioannou, P.A. (1987), Instability Analysis and Robust Adaptive Control of Robotic Manipulators, University of California, Electrical Engineering-Systems, Report 87-09-1.
F i g u r e 1.

Rohrs, C., Athans, M., and Valavani, L. (1985), Robustness of Continuous Time Adaptive Control Algorithms in the Presence of Unmodeled Dynamics, IEEE Trans. Aut. Coiiirol, AC-30, No. 9 Sadegh, N., and Horowitz, R., (1987), Stability Analysis of an Adaptive Controller for Robotic Manipulators, IEEE INi. Cor$ on Robotics and Autoniatron, Raleigh, NC. Singh, S.N., (1985), Adaptive Model Following Control of Nonlinear Robotic Systems, IEEE Trans. Auio. Conir., AC-30, No. 11. Slotine, J.-J. E., and Li. W., (1986), On the Adaptive Control of Robot Manipulators, ASME Winter Annual Meering, Anaheim, CA. Slotine, I.-J.E., aiid Li, W., (1987a), On the Adaptive Control of Robot Manipulators, Ini. J . of Robotics Research, Vol. 6, No. 3, pp. 49-59. Slotine, J.-J., E., and Li (1987b), Adaptive Robot Control: A New Perspective, Proc. IEEE Corzf. on Decision and Control, Los Angeles. Spong, M.W. (1987), Modeling and Control of Elastic Joint Manipulators, J. Dyn. Sys., Meas. and Conirol, Vol. 109, 310-319. Spong, M.W. (1988), Adaptive Control of Flexible Joint Robots, Systems and Control Leiiers, submitted for publication. Spong, M.W., and Ortega, R. (1988), On Adaptive Inverse Dynamics Control of Rigid Robots, IEEE J. of Robotics arzd Automation, submitted for publication. Spong, M.W., and Vidyasagar, M. (1987), Robust Linear Compensator Design for Noiiliiiear Robotic Control, IEEE J . of Robotics arid Autorizaiion, RA-3, No. 4, 345-351.
arid Corzrrol, Spong, M.W., aiid Vidyasagar, M. (198S), Robot Dyrza~nics John Wiley and Sons, Inc., New York.

F i g u r e 2.

Sweet, L. M. and M. C. Good, (1984), Redefinition of the robot motion control problem: effects of plant dynamics, drive system constraints, and user requirements, Proc. 23rd IEEE CDC, Las Vegas, NV.

1584