Experimental investigation on adaptive robust controller designs applied to a free-floating space manipulator
Tatiana F.P.A.T. Pazelli
a, Marco H. Terra
a, Adriano A.G. Siqueira
b,aDepartment of Electrical Engineering, University of Sao Paulo at S~ ao Carlos, Brazil~
bDepartment of Mechanical Engineering, University of S~ao Paulo at S~ao Carlos, Av. Trabalhador S~ao-carlense, 400, CEP 13566-590, S~ao Carlos, SP, Brazil
a r t i c l e i n f o
Article history:
Received 13 October 2009 Accepted 18 December 2010 Available online 3 March 2011 Keywords:
Space robotics H1control Adaptive control Neural networks Fuzzy systems
a b s t r a c t
This paper aims to formulate and investigate the application of various nonlinearH1control methods to a free-floating space manipulator subject to parametric uncertainties and external disturbances.
From a tutorial perspective, a model-based approach and adaptive procedures based on linear parametrization, neural networks and fuzzy systems are covered by this work. A comparative study is conducted based on experimental implementations performed with an actual underactuated fixed- base planar manipulator which is, following the DEM concept, dynamically equivalent to a free-floating space manipulator.
&2011 Elsevier Ltd. All rights reserved.
1. Introduction
Considerable research efforts have been directed to some primary functions of manipulator robots for in-space operations, such as assembly, inspection and maintenance. A number of dynamic and control problems are unique to this area due to the distinctive and complex dynamics found in many of these applications.
The main feature of space robots is the dynamic coupling between the spacecraft and the robotic arm. It causes the coordinated rotation of the main body with the motions of the arm.Dubowsky and Papadopoulos (1993)identified representa- tive types of space robotic systems considering some of their specific planning and control problems: free-flying space manip- ulators and free-floating space manipulators.
In a free-flying space manipulator the spacecraft attitude is actively controlled by reaction wheels and/or jets. However, the use of these mechanisms increases the consumption of electrical power and fuel, also adding more weight, complexity and dis- turbances to the system. The consumption of relatively large amounts of fuel can limit the system’s useful life in space. Thus, to deal with this limitation, trajectory planning control approaches for this type of system are formulated in order to optimize energy consumption (Sakawa, 1999; Torres and Dubowsky, 1992).
Differently, free-floating space manipulators (SM) are systems that allow the spacecraft to move freely in response to manipulator
motions with the purpose of conserving energy (Papadopoulos and Dubowsky, 1991). Considering this configuration, coordinated spacecraft/manipulator motion control has been addressed in case of redundant manipulators (Caccavale and Siciliano, 2001;
Dubowsky and Torres, 1991). Trajectory planning algorithms have also been developed in order to minimize the reaction of the free- floating base while the manipulator task is executed (Huang and Xu, 2006; Liu et al., 2009; Papadopoulos et al., 2005; Torres and Dubowsky, 1992; Tortopidis and Papadopoulos, 2006).
Although each of these types of space robotic systems demands a specific controller design, the combination of both can be employed during different phases of a mission. This paper focuses on the optimization of the motion control of a free- floating space manipulator. Online trajectory tracking is evaluated considering a predefined trajectory.
In virtue of the complexity of these systems and the environ- ment where they operate, the choice of a control system to this class of manipulators assumes an important role to deal with parametric uncertainties and external disturbances. When differ- ent payloads are being handled during different tasks, mass and inertial characteristics are difficult to be estimated precisely. The hostile work environment can also generate model uncertainties due to wear of mechanic and electronic components. Moreover, most space manipulators are designed to be light weight and low power systems, what is consistent with the zero gravity environ- ment and energy efficiency concerns. As a result, joint friction, damping, and other system nonlinear uncertainties are much more expressive in space robots than in industrial robots. There- fore, robustness is a significant challenge to the motion control of space manipulators. Two approaches to control uncertain systems Contents lists available atScienceDirect
journal homepage:www.elsevier.com/locate/conengprac
Control Engineering Practice
0967-0661/$ - see front matter&2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2010.12.011
Corresponding author. Tel.: + 55 16 3373 9398; fax: + 55 16 3373 9402.
E-mail address:siqueira@sc.usp.br (A.A.G. Siqueira).
subject to external disturbances are commonly used in controller designs: adaptive control and robust control.
Adaptive controllers estimate and compensate the uncertain dynamics of a system, (Craig, 1988). The adaptive procedure is generally based on the linear parametrization property, which demands a precise knowledge of the model structure. It considers that the unknown parameters are constant or vary slowly.
Besides, unmodeled dynamics are usually present and their effects decrease the performance of this approach. This method has been addressed to the motion control of free-floating space manipulators in several works in the literature (e.g.,Gu and Xu, 1993; Parlaktuna and Ozkan, 2004; Woerkom et al., 1996).
However, most of them require the measurement of orientation, velocity and acceleration of the free-floating base, which are difficult measures to be obtained in practice.
Another line of research on adaptive control is defined by intelligent systems, which exhibit interesting abilities of learning and interpretation. Approaches based on fuzzy and neural network systems are able to learn and approximate real-world concepts, building a knowledge base that may be interpreted and modified by the user. Moreover, these approaches have high potential to provide mechanisms for building systems subject to rapidly varying
unknown parameters. Thus, intelligent systems have been success- fully applied in the literature to universally approximate mathema- tical models of dynamic systems, see for instance (Hornik et al., 1989;
Jang, 1993; Narendra and Parthasarathy, 1990; Polycarpou, 1996;
Takagi and Sugeno, 1985). Within the scope of this work, applications of intelligent systems to identify and control free-floating space manipulators can be found in Sanner and Vance (1995); Taveira et al. (2006); Guo and Chen (2008); Huang and Chen (2008) andPazelli et al. (2008).
On the other hand, robust controllers have been designed to deal with parametric uncertainties and external disturbances. In the case of space robots, disturbances may be generated, for example, by spacecraft movements, microgravity effects and sensor and actuator noises due to extremes of temperature and glare. Impact effects during the docking or rendezvous process are also representative disturbances for this kind of robot. Sliding mode control is one of the main approaches that deals with disturbance effects. However, this method is usually accompanied by a phenomenon called chattering, which can increase energy consumption and damage the actuators. In addition, due to high frequency content of chattering, it can easily stimulate flexible modes which in turn may cause instability. This issue is extensively worked throughout the literature, and different Nomenclature
SM parameters and coordinates
miu mass of linki Ciu mass center of linki Iiu inertia tensor of linki
Jiu joint connecting linkði1Þto linki Li vector connectingJiutoCiu
Ri vector connectingCiutoJuiþ1 Miu total mass
yiu rotation of linkiaround jointJui
ð
j
u,yu,cuÞEuler angles (Z-Y-Z) w.r.t. the spacecraft attitude qu generalized coordinatesDEM parameters and coordinates
mi mass of linki Ci mass center of linki Ii inertia tensor of linki
Ji joint connecting link (i1) to linki Wi vector connectingJitoJi+ 1
lci vector connectingJitoCi Mt total mass
yi rotation of linkiaround jointJi
ð
j
,y,cÞ Euler angles (Z-Y-Z) w.r.t. the base attitude (J1) q generalized coordinatesDEM model
M(q) symmetric positive definite inertia matrix Cðq,qÞ_ matrix of Coriolis and centrifugal forces
t
torque vector acting upon the joints of the DEM b index w.r.t. the passive spherical joint (base) m index w.r.t. the active joints (manipulator)Problem formulation
qmd desired reference trajectory for manipulator joints x~m state tracking error
NonlinearH1Control
g
desired disturbance attenuation level u control variableQ symmetric positive definite weighting matrix w.r.t.
x~m
R symmetric positive definite weighting matrix w.r.t.u
Adaptive nonlinearH1control
X matrix of known functions F vector with uncertain components
Z symmetric positive definite weighting matrix w.r.t.F_
Adaptive neural network nonlinearH1control
Hðxe,FÞ set ofnneural networks
Fk vector of adjustable weights in the output layer wk vector of constant weights in the input layer bk vector of constant biases in the hidden layer G¼tanh activation function for the neurons in the
hidden layer
Adaptive fuzzy nonlinearH1control
Hðxe,FÞ set ofnfuzzy systems
Fk vector of adjustable parameters in the output func- tional consequences
uki fuzzified input variables Aij linguistic variables
yl output functional consequences
m
l d.o.f. ofl-th ruleprocedures have already been applied to diminish this effect.Feng et al. (2008)proposed a robust sliding mode controller with a radial basis function neural network to eliminate the chattering phenom- enon in the tracking control problem of a free-floating space robot.Guo and Chen (2006)have presented a robust control scheme for a dual-arm space robot system based on the augmentation approach proposed by Gu and Xu (1993). Zhongyi et al. (2008) employed a disturbance observer in each joint of the arm to estimate plant uncertainties, unmodeled dynamics and external disturbances within the execution of a free-floating space manipulator task. In this scheme, a low pass filter plays a significant role in ensuring the robustness and disturbance suppression performance of the system.
InPathak et al. (2008), robust trajectory control of free-floating space robots is performed by a set of linear time invariant controllers coupled to the robotic manipulator through a set of high feed-forward gains and suitably defined feedback paths. In this reference, the robustness of the control method is guaranteed in virtue of the controller not demand the knowledge of the manipulator parameters.
A different method of robust control, based on theH1criterion, has been widely used among robotic applications for guaranteeing good disturbance rejection properties. TheH1norm is a measure of the robustness applied by the control system to the plant. It defines the level of attenuation in the input/output relationship between the disturbance and the controlled output. Applications of this approach include the control of position and attitude of satellites and space- crafts (Chen et al., 2000; Dalsmo and Egeland, 1997; Kang, 1995;
Skullestad and Gilbert, 2000; Yang and Sun, 2002). Concerning the motion control of free-floating manipulators, however, works that have resorted to this method are difficult to find.
Considering the presented scenario, this paper aims to inves- tigate the application of nonlinear H1 controllers to a free- floating space manipulator subject to parametric uncertainties and external disturbances. To conduct a comparative study, this work covers the following methods of robust control based on the H1 criterion: (1) Nonlinear H1 control based on game theory;
(2) Adaptive nonlinearH1 control; (3) Adaptive neural network nonlinear H1 control; and (4) Adaptive fuzzy nonlinear H1
control. To the best of the authors’ knowledge, a comparison among adaptive nonlinear H1 controllers has not been made throughout the literature.
Method (1), derived fromSiqueira and Terra (2004), and method (2), based on linear parametrization, demand a complete knowledge of the nominal model structure. The intelligent procedures, (3) and (4), are proposed in two different approaches. The first one applies the intelligent system to learn the dynamic behavior of the robotic system, which is considered totally unknown. The second approach considers a well-defined nominal model structure for the arm and the intelligent system is applied to estimate only the behavior of uncertain dynamics. It must be noted that these controllers do not demand any kinematic or dynamic data from the spacecraft. TheH1
criterion is applied to all the proposed techniques to attenuate the effect of estimation errors and external disturbances.
An important issue of this comparative study is that the analysis of controllers requires accurate simulations of plant behavior. However, it is difficult to recreate space conditions on Earth. Thus, a modeling approach that relates space and terrestrial systems is opportune. An analytical method of modeling free-floating space robots, called Virtual Manipulator (VM), was proposed in Vafa and Dubowsky (1990). The VM is an inertial fixed-base robot whose first joint is a passive spherical one, representing the free-floating nature of the space manipulator. However, only kinematic equivalence is consid- ered in this representation. Based on the VM concept,Liang et al.
(1997)mapped a SM to a conventional fixed-base manipulator and showed that both kinematic and dynamic properties of the space manipulator system are preserved in this mapping. This manipulator is called Dynamically Equivalent Manipulator (DEM). The DEM goes
beyond the VM concept since it models the free-floating space manipulator system both kinematically and dynamically. Thus, the DEM can be physically built using a conventional manipulator system. It allows to experimentally study the dynamic performance and task execution of a space manipulator system without having to resort to complex experimental set-ups to simulate the space environment. Nevertheless, experimental applications using this resource are not found in literature.
Therefore, in order to validate and compare the proposed control procedures in this paper, experimental implementations are performed with an actual underactuated fixed-base planar manipulator. Based on the DEM concept, this robot is dynamically equivalent to a free-floating space manipulator. A qualitative analysis of the results is developed concerning the ability of attenuating the effects of the disturbance. In addition, a quanti- tative description of the results is presented in the form of performance indexes which express the error of trajectory track- ing and the energy consumption.
In summary, the contribution of this paper consists of an experimental investigation on the motion control of a free- floating space manipulator subject to parametric uncertainties and external disturbances, performed by different methods of adaptive nonlinearH1controllers.
This paper is organized as follows: the kinematic and dynamic equivalences between the free-floating space manipulator and a fixed-base manipulator are presented in Section 2; the solutions for the nonlinear H1 control problems based on the linear parame- trization property of the model, neural networks and fuzzy systems are presented in Section 3; and, finally, experimental results for a two-link free-floating space manipulator are presented in Section 4.
2. Model description and problem formulation
2.1. Free-floating space manipulator mapped by a dynamically equivalent fixed-base manipulator
Consider ann-link serial-chain rigid manipulator mounted on a free-floating base and that no external forces and torques are applied on this system. Consider also the Dynamically Equivalent Manipulator (DEM) approach (Liang et al., 1997). The DEM is a (n+ 1)-link fixed-base manipulator with its first joint being a passive spherical one and whose model is both kinematically and dynamically equivalent to the SM dynamics. Since it is a conventional manipulator, it can be physically built and experi- mentally used to study control algorithms for space manipulators.
Fig. 1 shows the representation and parameter notation for both SM and DEM. Let the SM parameters be identified by apostrophe (0), the links of the manipulators are numbered from 2 ton+1 andJiis the joint connecting the (i1)-th link andi-th link,yiis the rotation of thei-th link around jointJi, and the Z-Y-Z Euler anglesð
j
,y,cÞrepresent the SM base attitude and the DEMFig. 1.The space manipulator and its corresponding DEM (Liang et al., 1997).
first passive joint orientation. The Z-Y-Z convention of the Euler angles is used in this paper as in Liang et al. (1997). This convention represents a rotation of the XYZ-system about theZ- axis by
j
; a rotation of the XuYuZu-system about theXu-axis byy, and a rotation of the X00Y00Z00-system about the Z00-axis byc.LetCibe the center of mass of thei-th link,Libe the vector connectingJui andCui,Ribe the vector connectingCuiandJuiþ1,lci be the vector connectingJiandCi, andWibe the vector connecting JiandJi+ 1. Considering that the DEM operates in the absence of gravity and that its base is located at the center of mass of the SM, the kinematic and dynamic parameters of the DEM can be described from the SM parameters as
mi¼ M2tmui Pi1
k¼1mukPi k¼1muk
,
I1¼Iu1, Ii¼Iui,
W1¼R1
Mt
Xi
k¼1
muk,
Wi¼ Ri
Mt
Xi
k¼1
mukþ Li
Mt
Xi1
k¼1
muk,
lc1¼0,
lci¼ Li
Mt
Xi1
k¼1
muk, ð1Þ
wherei¼2,y,n+1 andMtis the total mass of the SM. Observe that the mass of the passive joint, m1, is not defined by the equivalence properties.
From Lagrange theory, dynamic equations of the DEM are given by a set of nonlinear coupled differential equations as
MðqÞq€þCðq,qÞ_ q_¼
t
, ð2Þwhereq¼ ½
j
y c y2 ynþ1TARðnþ3Þare generalized coordi- nates,MðqÞARðnþ3Þðnþ3Þis the symmetric positive definite iner- tia matrix, Cðq,qÞ_ ARðnþ3Þðnþ3Þ is the matrix of Coriolis and centrifugal forces, andt
¼ ½0 0 0t
2t
nþ1TARðnþ3Þ is the torque vector acting upon the joints of the DEM.From a practical point of view, the SM must handle different payloads whose mass and inertial characteristics can be difficult to obtain. Also, the dynamics of joint friction, damping, vibration or other nonlinear uncertainties are not easy to model or measure, producing effects that cannot be neglected. Moreover, the hostile work environment and the floating base are sources of disturbances which can also generate trajectory errors in the SM.
Due to this fact, a finite energy torque disturbance may be defined by
t
dand modeling uncertainties can be introduced dividing the parameter matrices M(q) and Cðq,qÞ_ into a nominal and a perturbed part:MðqÞ ¼M0ðqÞ þDMðqÞ,
Cðq,qÞ ¼_ C0ðq,qÞ þ_ DCðq,qÞ,_
where M0(q) and C0ðq,qÞ_ are nominal matrices and DMðqÞ and DCðq,qÞ_ are the model uncertainties.
2.2. Problem formulation
Since a free-floating space manipulator is being considered, it is assumed that only the active joints of the DEM are controlled, with the passive spherical joint not locked. In this case, the
passive joint dynamics intervene with the control of the manip- ulator active joints.
Letqbe partitioned asq¼ ½qTb qTmT, where the indexesbandm represent the passive spherical joint (base) and the active joints (manipulator), respectively. Defined¼ ½dTb dTmTas a vector repre- senting the sum of parametric uncertainties of the system and
t
d¼ ½t
Tdb
t
TdmT as the finite energy external disturbance also introduced. Eq. (2) can be rewritten as0 tm
" # þ db
dm
" # þ tdb
tdm
" #
¼ Mbb Mbm
Mmb Mmm
" # q€b q€m
" #
þ Cbb Cbm
Cmb Cmm
" # q_b
q_m
" # , ð3Þ whereMbbðqmÞAR33,MbmðqmÞAR3n,MmbðqmÞARn3,MmmðqmÞA Rnn,Cbbðqm,qÞ_ AR33,Cbmðqm,qÞ_ AR3n,Cmbðqm,qÞ_ ARn3,Cmmðqm, q_mÞARnn are nominal matrices and
t
mARn. For simplicity of notation, the 0 index referring to the nominal system was suppressed. It must also be noted that (3) is formulated so that the inertia matrix related to the controlled joints, Mmm(qm), is symmetric positive definite and the matrixNmmðqm,q_mÞ ¼M_mmðqm, q_mÞ2Cmmðqm,q_mÞis skew-symmetric.Remark 1. The controller is designed for the SM considering the DEM model.
LetqdmARnandq_dmARnbe the desired reference trajectory and the corresponding velocity for the controlled joints, respectively.
The state tracking error is defined as
x~m¼ q_mq_dm qmqdm
" #
¼ q_~m
q~m
" #
: ð4Þ
The variablesqmd
,q_dm, andq€dm(desired acceleration) are assumed to be within the physical and kinematic limits of the control system and no reference trajectory exists for the base.
In order to minimize x~m acting with minimal torque and energy consumption, a control input is chosen according to Johansson (1990)as
u¼ ½Mmm Cmm z_~1
z~1
" #
¼MmmT1x_~mþCmmT1x~m, ð5Þ
withz~1andT1being introduced by the transformation
z~¼ z~1
z~2
" #
¼T0x~m¼ T1
T2
" #
x~m¼ T11 T12
0 I
q_~m
q~m
" #
, ð6Þ
whereT11,T12ARnnare constant matrices to be determined. The defined control input (5) includes selective torques that affect only the kinetic energy, the reference trajectories (4) and a state space transformation (6) that eliminates redundancies among position and velocity coordinates while keeping full state space order. Consequences of the choice of this transformation are that the synthesis of the nonlinearH1control depends on an algebraic Riccati equation, which is easily solved, and the resulting gain will be constant.
From (3)–(6), state space representation of the DEM is given by x_~m¼ATðqm,q_mÞ~xmþBTðqmÞT11ð
t
mFðxemÞEðxebÞ þdmþt
dmÞ, ð7Þ whereATðqm,q_mÞ ¼T01 M1mmCmm 0 T111 T111T12
" #
T0,
BTðqmÞ ¼T01 M1mm 0
" # ,
FðxemÞ ¼Mmmðq€dmT111T12q_~mÞ þCmmðq_dmT111T12q~mÞ,
and
EðxebÞ ¼Mmbq€bþCmbq_b,
withxem¼ ½qTm q_Tm ðqdmÞT ðq_dmÞT ðq€dmÞTTandxeb¼ ½_qTb q€TbT.
2.3. NonlinearH1control
Given a desired disturbance attenuation level of
g
40, the state feedback nonlinearH1 control method aims to attenuate the disturbance in the system through a control law of the form u¼Kð~xmÞx~min order to satisfy the performance criterionuðÞminAL2
0amaxoðÞAL2
R1
0 1
2x~TmðtÞQx~mðtÞ þ12uTðtÞRuðtÞ
dt R1
0 1
2
o
TðtÞo
ðtÞ
dt r
g
2, ð8ÞwhereQandRare symmetric positive definite weighting matrices defined by the designer,x~mð0Þ ¼0 and
o
¼T11ðdmþt
dmÞrefers to the disturbance term in (7).Following the game theory, a solution of thisminimaxproblem is given byChen et al. (1994), in a simplified form, in terms of the algebraic equation
0 K K 0
TT0B R11
g
2I
BTT0þQ¼0, ð9Þ
withB¼ ½I 0T. This simplification in the synthesis of the non- linearH1controller is the main consequence of the transforma- tion (6). The resulting gain will be constant in a similar way to the state feedback linearization approaches.
Hence, for the proposed application, the H1 control pro- blem (8) subject to (7) has an optimal solution
u¼ R1BTT0x~m, ð10Þ
if there are matricesK40 and a non-singularT0solutions of (9), withRo
g
2I.Considering that the matrixR1is obtained through a Cholesky decomposition of theR-dependent term in (9)
RT1R1¼ R11
g
2I1
, andQis factorized as
Q¼ Q1TQ1 Q12
Q12T Q2TQ2
" #
,
the solution of (9) is given by
T0¼ RT1Q1 RT1Q2
0 I
" #
and K¼1
2ðQ1TQ2Q2TQ1Þ1
2ðQ21T þQ12Þ:
Thus, from (7), the applied torques which guarantee the desired H1performance (8) can be computed by
t
m¼T111uþFðxemÞ þEðxebÞ: ð11ÞRemark 2.This nonlinearH1 control method assumes that the model structure is completely known and represents parameter uncertainties as internal disturbances, modeling them in the same way as external disturbances.
3. Adaptive robust controller design
The adaptive control designs presented in the following apply different learning methods to estimate uncertain parameters and also the behavior of unmodeled dynamics. TheH1control law is applied to attenuate the effects of estimation errors and external disturbances.
Hence, given a desired level of attenuation
g
40 and matrices Q¼QT40,R¼RT40,P0¼PT040 and Z0¼Z0T40, the adaptivenonlinear H1 control problem is defined for the proposed application by the following performance criterion
Z T 0
ðx~TmQx~mþuTRuÞdtrx~Tmð0ÞP0x~mð0Þ þF~Tð0ÞZ0F~ð0Þ þg2 Z T
0
ðoToÞdt, ð12Þ whereF~ ¼FFdenotes the parameter estimation error.
3.1. Adaptive nonlinearH1control
Adaptive nonlinear H1 control presented in this section assumes that the model structure of the system is well-defined but its formulation is based on uncertain (or unknown) para- meters which determineM(q) andCðq,qÞ._
Assume the linear parametrization property of the term FðxemÞdmin (7):
FðxemÞdmHðxe,FÞ ¼XF, ð13Þ where xe is the input vector of Xðqm,q_m,q_dmT11T12q~m,q€dm T111T12q_~mÞ, being X a ðnpÞ regression matrix of known func- tions, andFis ap-dimensional vector with uncertain components depending on manipulator parameters.
Let the optimal approximation parameters vector be F¼arg min
FAOF
xmaxeAOxeJHðxe,FÞðFðxemÞdmÞJ2,
whereJJ2is the Euclidean norm.
Define
t
¼t
mEðxebÞ and the modified error Eq. (7) may be rewritten asx_~m¼ATðqm,q_mÞ~xmþBTðqmÞT11ð
t
FðxemÞ þdmþt
dmþHðxe,FÞHðxe,FÞÞ
¼ATðqm,q_mÞx~mþBTðqmÞT11ð
t
Hðxe,FÞÞ þBTðqmÞT11ðHðxe,FÞFðxemÞ þdmþt
dmÞ¼ATðqm,q_mÞx~mþBTðqmÞuþBTðqmÞ
o
ð14Þwith
u¼T11ð
t
Hðxe,FÞÞ, ð15Þo
¼T11ðHðxe,FÞFðxemÞ þdmþt
dmÞ, ð16Þ whereo
refers to the estimation error from the adaptive system and external disturbances. Using u¼u the control law provided by the nonlinearH1controller in (10),t
can be computed byt
¼Hðxe,FÞ þT111u: ð17ÞConsidering the solvability analysis developed by Chen et al.
(1997), a dynamic state feedback controller given by
F_ ¼
a
ðt,x~mÞ ¼ Z1XTT11BTT0x~m, ð18Þt
¼t
ðt,F,x~mÞ ¼XFT111R1BTT0x~m, ð19Þ with Z¼ZT40, is solution to the adaptive nonlinear H1control problem subject to (14) and satisfies (12) for any initial condition.
Factoring out q€b in the first line of (3) – disregarding uncer- tainties and disturbances – and substituting the result back on its second line, leads to
M¼MmmMmbM1bbMbm,
h¼ ðCmbMmbM1bbCbbÞq_bþ ðCmmMmbM1bbCbmÞq_m:
The relation between
t
and the torques applied upon the manipulator joints,t
m, is given byq€m¼ M^1mmðC^mmq_m
t
Þ, ð20Þt
m¼M^q€mþ^h, ð21Þ
whereM^mm,C^mm,M^ andh^ are the matricesMmm,Cmm,Mandh computed with the estimated valuesF.
Remark 3. This solution uses only the velocity of the free-floating base,q_b. If it was tried to linearly parametrize the complete term ðFðxemÞ þEðxebÞdmÞ in (7), the solution would also need the acceleration of the free-floating base q€b which is not easy to obtain in practice.
3.2. Adaptive neural network nonlinearH1control
Define a set ofnneural networksHkðxe,FkÞ,k¼1,y,n, where xe is the input vector andFk are the adjustable weights in the output layers. The single-output neural networks are of the form
Hkðxe,FkÞ ¼ Xpk
i¼1
fkiG Xqk
j¼1
wkijxejþbki 0
@
1
A¼xTkFk, ð22Þ
whereqkis the size of vectorxeandpkis the number of neurons in the hidden layer. The weightswijkand the biasesbikfor 1rirpk, 1rjrqkand 1rkrnare assumed to be constant and specified by the designer. Thus, the adjustment of neural networks is performed only by updating the vectors Fk. The activation function for the neurons in the hidden layer is chosen to be Gð:Þ ¼tanhð:Þ.Fig. 2illustrates the structure ofHk.
The complete neural network is denoted by
Hðxe,FÞ ¼
H1ðxe,F1Þ H2ðxe,F2Þ
^ Hnðxe,FnÞ 2
66 66 4
3 77 77 5¼
xT1 0 . . . 0
0 xT2 ^ 0
^ ^ & ^
0 0 . . . xTn
2 66 66 64
3 77 77 75
F1
F2
^ Fn
2 66 64
3 77
75¼XF, ð23Þ
with
xTk¼ tanh Xqk
j¼1
wk1jxejþbk1 0
@
1
A tanh Xqk
j¼1
wkp
kjxejþbkp
k
0
@
1 A 2
4
3 5,
Fk¼ ½fk1 fkpkT:
Consider a first approach where the term FðxemÞ þEðxebÞdm
in (7) is completely unknown regarding its structure and para- meter values. The neural network defined in (23) is applied to learn the dynamic behavior of the robotic system:
FðxemÞ þEðxebÞdmHðxe,FÞ ¼XF, ð24Þ where the input vectorxeshould be defined as
xe¼ ½qTm q_Tm q_Tb q€Tb ðqdmÞT ðq_dmÞT ðq€dmÞTT: ð25Þ Note that this definition ofxetakes into account the values of q_bandq€b, which was inspired in the mathematical model of the space manipulator aforementioned. However, these values are not easy to be measured in practice. Considering that a neural
network-based approach is in general used when it is not possible to supply all the variables values to the system model, the vector xecan be redefined as
xe¼_ xem¼ ½qTm q_Tm ðqdmÞT ð_qdmÞT ðq€dmÞTT, ð26Þ avoiding the necessity of any data from the free-floating base.
Experimental results described in Section 4 show the feasibility of this assumption. Defining the following optimization problem F¼arg min
FAOFmax
xeAOxeJHðxe,FÞðFðxemÞ þEðxebÞdmÞJ2, the modified error Eq. (7) may be rewritten as
x_~m¼ATðqm,q_mÞ~xmþBTðqmÞuþBTðqmÞ
o
ð27Þ withu¼T11ð
t
mHðxe,FÞÞ, ð28Þo
¼T11ðHðxe,FÞFðxemÞEðxebÞ þdmþt
dmÞ, ð29Þ whereo
refers to the estimation error from the neural network system and external disturbances. Considering the results achieved by Chang and Chen (1997), a dynamic state feedback controller given byF_ ¼
a
ðt,x~mÞ ¼ ZTXTT11BTT0x~m, ð30Þt
m¼t
mðt,F,x~mÞ ¼XFT111R1BTT0x~m, ð31Þ withZ¼ZT40, is solution to the adaptive neural network non- linearH1 control problem subject to (27) and satisfies (12) for any initial condition.Consider now a second approach where the model structure and nominal values for the termFðxemÞare well-defined and available for the controller. In this case, the neural network is applied to estimate only the behavior of parametric uncertainties and spacecraft dynamics (considered as an unmodeled disturbance):
EðxebÞdmHðxe,FÞ ¼XF: ð32Þ Similarly, xe¼_ xem, the optimal approximation parameters vector is given by
F¼arg min FAOFmax
xeAOxeJHðxe,FÞðEðxebÞdmÞJ2,
and the modified error Eq. (27) may be rewritten as
x_~m¼ATðqm,q_mÞ~xmþBTðqmÞuþBTðqmÞ
o
ð33Þ withu¼T11ð
t
mFðxemÞHðxe,FÞÞ, ð34Þo
¼T11ðHðxe,FÞEðxebÞ þdmþt
dmÞ, ð35Þ whereo
refers to the estimation error from the neural network system and external disturbances. This is an original proposal that follows the stability analysis developed byChang and Chen (1997) and applied to (12). A dynamic state feedback controller chosen as F_ ¼a
ðt,x~mÞ ¼ ZTXTT11BTT0x~m, ð36Þt
m¼t
mðt,F,x~mÞ ¼FðxemÞ þXFT111R1BTT0x~m, ð37Þ withZ¼ZT40, is solution to the adaptive neural network non- linearH1control problem subject to the distinct formulation (33) and satisfies (12) for any initial condition.3.3. Adaptive fuzzy nonlinearH1control
The Adaptive Network-based Fuzzy Inference System (ANFIS) defined byJang (1993)introduces a training procedure for Takagi–
Sugeno (T-S) fuzzy inference systems. Considering an input/output dataset, ANFIS builds a fuzzy inference system whose membership functions parameters are adjusted by a hybrid learning rule which Gkpk
xe
Wk
Gk1
Gk2
Gk...
Φk
ξk
b1k b2k b...k
bpkk
kpk
k...
k2
k1
Hk(xe, Φk)
Fig. 2.Structure of a single-output neural network.
combines the gradient method and the least squares estimate. An adaptive network architecture, similar to neural networks structure, consists of nodes and directional links through which nodes are connected. Part or all of the nodes are adaptive, which means their outputs depend on the parameters pertaining to these nodes. The learning rule specifies how these parameters should be changed to minimize a prescribed error measure.
In this paper, an ANFIS was applied to an input/output dataset within an off-line procedure before the task execution. Thus, it provides a set of input membership functions to be used during the control action.
The T-S fuzzy model is characterized by a fuzzy rule base with functional consequences instead of fuzzy consequences, as
IFu1isA11andu2isA12yanduqk isA1qk, THENy1¼f10þf11u1þf12u2þ. . .þf1qkuqk,
^
IFu1isApk1andu2isApk2yanduqk isApkqk, THENypk¼fpk0þfpk1u1þfpk2u2þ. . .þfpkqkuqk,
whereAij,i¼1,y,pkandj¼1,y,qk, are linguistic variables referred to fuzzy sets defined on the input spacesU1,U2,y,Uqk;u1AU1, u2AU2,. . .,uqkAUqk are input variables values;pkis the number of fuzzy rules andqkis the size of an input vector.
The inferred output from the T-S method is crisp (hence, it does not demand a defuzzifier) and it is defined by the weighted average of outputsylfrom each linear subsystem implied as
Hkðxe,FkÞ ¼ Ppk
l¼1
m
lylPpk
l¼1
m
l¼ Ppk
l¼1
m
lðfl0þfl1u1þfl2u2þ. . .þflqkuqkÞ Ppkl¼1
m
l¼xTkFk, ð38Þ Table 2
DEM parameters.
Body mi(kg) Ii(kgm2) Wi(m) lci(m)
Link 1 1.932 0.0153 0.203 0
Link 2 0.850 0.0075 0.203 0.096
Link 3 0.625 0.0060 0.203 0.077
Fig. 3.UnderActuated Robot Manipulator.
Table 3
Selected weighting matrices.
g¼2 R Q1 Q2 Z
NonlinearH1 0:9 0
0 2:5
0:64 0
0 0:64
9 0
0 7
–
AdaptiveH1 0:9 0
0 2:5
0:64 0
0 0:64
9 0
0 7
10 0
0 10
Adaptive NeuralH1(1) 0:9 0
0 2:5
0:64 0
0 0:64
9 0
0 7
10 0
0 10
Adaptive NeuralH1(2) 0:9 0
0 2:5
4 0
0 4
9 0
0 7
100 0
0 100
Adaptive FuzzyH1(1) 0:9 0
0 2:5
1 0
0 1
9 0
0 7
0:8 0
0 0:8
Adaptive FuzzyH1(2) 0:9 0
0 2:5
4 0
0 4
25 0
0 7
20 0
0 20
0 0.5 1 1.5 2 2.5 3 3.5
−60
−40
−20 0 20 40 60 80
Time (s)
Joint position (º)
Base Joint 2 Joint 3 Reference
0 0.5 1 1.5 2 2.5 3 3.5
−30
−20
−10 0 10 20 30 40 50
Time (s)
Joint velocity (º/s)
Fig. 4.Joints and base positions and velocities—nonlinearH1control.
Table 1 SM parameters.
Body mui(kg) Iui(kgm2) Ri(m) Li(m)
Base 4.816 0.0153 0.253 0
Link 2 0.618 0.0075 0.118 0.120
Link 3 0.566 0.0060 0.126 0.085
where 1rkrn and
m
l is the degree of freedom of l-th rule, defined as the minimum among the grade of membership asso- ciated to the entries in the activated fuzzy sets by thel-th rule,m
l¼Al1ðu1Þ4Al2ðu2Þ4. . .4AlqkðuqkÞ: ð39Þ
Thus, a set of fuzzy inference systems based on the T-S method is defined as
Hðxe,FÞ ¼
H1ðxe,F1Þ H2ðxe,F2Þ
^ Hnðxe,FnÞ 2
66 66 4
3 77 77 5¼
xT1 0 . . . 0
0 xT2 ^ 0
^ ^ & ^
0 0 . . . xTn
2 66 66 64
3 77 77 75
F1
F2
^ Fn
2 66 64
3 77
75¼XF, ð40Þ
with
xTk¼ 1 Ppk
l¼1
m
kl½m
k1m
k1xTem
k2m
k2xTem
kpkm
kpkxTe, Fk¼ ½fk10 fk11 . . . fk1qk fk20 fk21 . . . fk2q
k fkp
k0 fkp
k1 . . . fkp
kqkT, wherexeARqk is the input vector,Xis anðnpÞmatrix depen- dent on the input values and on the values of fuzzy rules degrees of freedom and the vector FARnðpkðqkþ1ÞÞ is a p-dimensional vector representing the adjustable parameters in the output functional consequences, withp ¼n(pk(qk+ 1)).
Consider a first approach where the term FðxemÞ þEðxebÞdm
in (7) is completely unknown regarding its structure and para- meter values. The T-S fuzzy system defined in (40) is applied to
learn the dynamic behavior of the robotic system:
FðxemÞ þEðxebÞdmHðxe,FÞ ¼XF: ð41Þ For the proposed application, the input vector is defined as xe¼x~m¼ ½_~qTm q~TmT and Að~xmÞ ¼ ½A1ð_~qmÞ A2ðq~mÞ comprises the fuzzy sets defined for the fuzzified inputs.
Observing the similar formulation in Section 3.2 and based on the analysis presented in Chang and Chen (1997) and Chang (2005), a dynamic state feedback controller given by
F_ ¼
a
ðt,x~mÞ ¼ ZTXTT11BTT0x~m, ð42Þt
m¼t
mðt,F,x~mÞ ¼XFT111R1BTT0x~m, ð43Þ with Z¼ZT40, satisfies (12) for any initial condition. It is solution to the adaptive fuzzy nonlinear H1 control problem subject to (27), where, in this case,o
refers to the estimation error from the T-S fuzzy system and external disturbances.The second approach proposed in Section 3.2 may also be considered for the fuzzy-based controller. In this case, the T-S fuzzy system is applied to estimate only the behavior of parametric uncertainties and spacecraft unmodeled dynamics:
EðxebÞdmHðxe,FÞ ¼XF: ð44Þ As in (36)–(37), a dynamic state feedback controller given by F_ ¼
a
ðt,x~mÞ ¼ ZTXTT11BTT0x~m, ð45Þt
m¼t
mðt,F,x~mÞ ¼FðxemÞ þXFT111R1BTT0x~m, ð46Þ with Z¼ZT40, satisfies (12) for any initial condition. It is solution to the adaptive fuzzy nonlinear H1 control problem subject to (33), where, in this case,o
refers to the estimation error from the T-S fuzzy system and external disturbances.Remark 4. Within both neural and fuzzy control methods, measured values for orientation, velocity and acceleration of the free-floating base are not necessary. The intelligent systems are applied to estimate the spacecraft dynamics as well as parametric uncertainties and unmodeled dynamics. It can be seen by the control laws defined by (30)–(31), (36)–(37), (42)–(43) and (45)–(46) that the spacecraft coordinatesqb,q_b and q€b are not included among the entries ofFðxemÞ,Hðxe,FÞandu.
4. Results
For validation and comparison purposes, the proposed adap- tiveH1 control solutions are applied to a free-floating, planar, two-link space manipulator system, whose nominal parameters are given in Table 1. The corresponding DEM is a fixed-base, three-link, planar manipulator whose first joint is configured as passive, that is,qm¼ ½q2 q3T are the joints to be controlled. The
0 0.5 1 1.5 2 2.5 3
−0.1
−0.05 0 0.05 0.1 0.15
Time (s)
Torque (N.m)
Joint 2 Joint 3
Fig. 5.Torques—nonlinearH1control.
0 0.5 1 1.5 2 2.5 3
−60
−40
−20 0 20 40 60 80
Time (s)
Joint position (º)
Base Joint 2 Joint 3 Reference
0 0.5 1 1.5 2 2.5 3
−30
−20
−10 0 10 20 30 40
Time (s)
Joint velocity (º/s)
Fig. 6.Joints and base positions and velocities—adaptive nonlinearH1control.