à áá à Ö n ? . 5 6
A Robust Hybrid Intelligent Position/ForceControl Scheme for Cooperative
Manipulators
Wail Gueaieb
School of Information Technology and EngineeringUniversity of Ottawa
Wail Gueaieb 1 of 28
à áá à Ö n ? . 5 6
Talk Overview
Ô Introduction
Ô Problem Statement
Ô Standard Adaptive Control of CR
Ô Adaptive Fuzzy Controllers
Ô A Hybrid Adaptive Control Scheme
Ô Conclusions
Ô Future Research Directions
Wail Gueaieb 2 of 28
à áá à Ö n ? . 5 6
Introduction
z Why Cooperative Robots?F More efficient handling of certain objects
ã Cardboardsã Large sheets of glassã Heavy and/or large objects, in general
F Certain tasks may be too complex for a single manipulator system
ã Space missionsã Underwater oil pipelines maintenance
Wail Gueaieb 3 of 28
à áá à Ö n ? . 5 6
Introduction (cont’d)
z Research ScopeAlthough cooperative robots usually consist of several modules, themain focus here is on:
O Position and force control, and
O External disturbance attenuation.
ManipulatorsJoints Joints
Wrists
Vision
VisionController
Device(s)Vision
Coordinator
Grippers
scope of this research
Modules within the
Controller
DisturbanceRejection
Coordinator
GripperController
GripperOverallMotion
Coordinator
ControllerCoordinationMotion and
Algorithmic Hardware Layer
Expert Layer
High
Low
Ban
dwid
th
Wail Gueaieb 4 of 28
à áá à Ö n ? . 5 6
Introduction (cont’d)
z Challenges of CR Control
O Not much research done in the control of strongly coupled CR.
O Control of strongly coupled CR is much more complex than that ofsingle robotic systems:
F Kinematic and dynamic coordination.
F Ubiquitous presence of uncertainties.
F Stricter stability criteria.
à Necessity to develop robust control approaches to keep up with theincreasingly demanding design requirements.
Wail Gueaieb 5 of 28
à áá à Ö n ? . 5 6
Introduction (cont’d)
z Brief History of CR Control
TypesMain
References Shortcomings
Non-adaptive Tarn et al. 87, 88, 92Bergerman et al. 98
F No control on internal forces.
F Model based: no uncertainties.
Adaptive
Hu et al. 93Vukobratovic et al. 98
Liu et al. 98Sun et al. 02
Szewczyk et al. 02
F No modeling uncertainties.
Soft computing Ge et al. 99
F No control on internal forces.
F Neural networkà too many parameters.
F Controller’s behavior not wellunderstood.
Wail Gueaieb 6 of 28
à áá à Ö n ? . 5 6
Problem Statement
O Consider two or morecooperative manipulatorsholding a common object.
E 2
Em
E1
Robot m
Robot 2
Robot 1
Object
O Control Objectives:
F Simultaneously
ã Track predefined object’s trajectory
(position and orientation.)
ã Make internal forces converge to de-sired values.
F In the presence of
ã Parametric (structured) uncertainties
(e.g., load’s mass and inertia.)
ã Modeling (unstructured) uncertainties(e.g., unknown time-varying externaldisturbances.)
Wail Gueaieb 7 of 28
à áá à Ö n ? . 5 6
CR Dynamics
In a CR system, the ith manipulator’s dynamics may be expressedas
τi = Di(qi)qi + Ci(qi, qi)qi + Gi(qi)− τdi− JT
φi(qi)fi
qi : joint coordinatesτi : joint torque/force applied by actuator (controller’s output)τdi
: disturbance vectorJφi
(qi) : Jacobian matrix from payload’s center of mass to qi
fi : internal force between end-effector and payloadDi(qi) : inertial matrix including payload’s inertia
Ci(qi, qi) : Coriolis and centrifugal matrix including payload’s termsGi(qi) : Gravitational vector including payload’s gravitational terms
Wail Gueaieb 8 of 28
à áá à Ö n ? . 5 6
Standard Adaptive Control of CR
O One of the most recent and efficient CACs was proposed by Liu et al.Control law of ith manipulator:
τi = Di(qi)qri+ Ci(qi, qi)qri
+ Gi(qi)−Ksisi − JT
φi(qi)(Kix + fdi
)
A : estimate of matrix A with parametric uncertainties onlyx = x− xd : position error of the payload’s center of massqri
= J+φi
(qi)(xd − γix) : the reference joint velocitysi = qi − qri
: residual error of the reference joint velocityJ+
φi(qi) : pseudo-inverse of Jφi
(qi)fdi
: desired internal forceKsi
and Ki : positive definite gain matrices
F Compensates for parametric uncertainties only.
F Assumes perfect knowledge of the working environment model.
à No compensation for modeling uncertainties nor for unstructuredexternal disturbances.
Wail Gueaieb 9 of 28
à áá à Ö n ? . 5 6
Standard Adaptive Control of CR (cont’d)
z Numerical ResultsF Two 3-DOF manipulators.
F Payload to follow an oblique line between the two manipulators.
F Internal forces lines of actions are not orthogonal to payload’s tra-jectory, with desired values fd1
= −fd2= 10 N.
F τd1= α(Γq1 + ρ(t) + λ), τd2
= −α(Γq2 + ρ(t) + λ)
Wail Gueaieb 10 of 28
à áá à Ö n ? . 5 6
Standard Adaptive Control of CR (cont’d)
y Experiment 1
F Parametric uncertainties only (payload’s mass).
F No modeling uncertainties and no unstructured external distur-bances (α = 0).
0 5 10 15 20 25 300.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
Time (s)
x (m
)
DesiredActual
0 5 10 15 20 25 300.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0.22
Time (s)
y (m
)
DesiredActual
0 5 10 15 20 25 305
10
15
20
25
30
35
Time (s)
ψ (d
eg)
DesiredActual
0 5 10 15 20 25 30−18
−16
−14
−12
−10
−8
−6
−4
−2
Time (s)
Inte
rnal
For
ce f 1 (N
)DesiredActual
Wail Gueaieb 11 of 28
à áá à Ö n ? . 5 6
Standard Adaptive Control of CR (cont’d)
y Experiment 2
F Modeling uncertainties are introduced in the form of unknown time-varying external disturbances (intensity level α = 1)
0 5 10 15 20 25 300.35
0.4
0.45
0.5
0.55
0.6
0.65
0.7
Time (s)
x (m
)
DesiredActual
0 5 10 15 20 25 300.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0.22
Time (s)
y (m
)
DesiredActual
0 5 10 15 20 25 305
10
15
20
25
30
35
40
45
50
Time (s)
ψ (d
eg)
DesiredActual
0 5 10 15 20 25 30−18
−16
−14
−12
−10
−8
−6
−4
−2
Time (s)
Inte
rnal
For
ce f 1 (N
)
DesiredActual
Wail Gueaieb 12 of 28
à áá à Ö n ? . 5 6
Soft Computing Based Controllers
O CR usually have very complex dynamics.à deriving a precise model is extremely difficult.
O Soft computing tools do not require a precise dynamics model.
O Main focus here: fuzzy logic based controllers (FLCs.)
F Rule-based expert systems: use of human-like linguistic variables,values, and simple if-then rules.
F Powerful in representing human knowledge.
Wail Gueaieb 13 of 28
à áá à Ö n ? . 5 6
Soft Computing Based Controllers (cont’d)
Working EnvironmentSensors Plant
CrispInput
CrispOutput
Fuzzy Fuzzy
KnowledgeBase
Decision-Making Logic(Decision Rules)
Fuzzification Defuzzification
Fuzzy Logic Controller
Wail Gueaieb 14 of 28
à áá à Ö n ? . 5 6
Soft Computing Based Controllers (cont’d)
z Merits of FLCs
å No need for a precise model.
å Robustness: tolerate noise and time-varying parameters in theplant’s dynamics.
å Generic: can be transferred from one platform to another withminor modifications.
z Drawbacks of FLCs
æ heavily dependent on human expertise.
æ Lack of efficient and systematic online adaptation mechanism toadapt to varying working conditions.
Wail Gueaieb 15 of 28
à áá à Ö n ? . 5 6
Adaptive Fuzzy Controllers
O Adaptive fuzzy controllers (AFCs) compensate for the shortcomings ofstatic FLCs while inheriting their strengths.
F Adaptation: ability to learn plant’s dynamics online.
F Higher robustness than CACs in the face of parametric and model-ing uncertainties.
O Adaptive fuzzy controller’s jth output:
yj =
L∑l=1
y(l)j ξl(x) = ΘT
j ξ(x)
ΘTj = (y
(1)j , . . . , y
(L)j ) , ξT (x) = (ξ1(x), . . . , ξL(x))
ξl(x) =
∏ni=1 µ
A(l)i(xi)∑L
k=1
( ∏ni=1 µ
A(k)i
(xi)), l = 1, . . . , L.
Wail Gueaieb 16 of 28
à áá à Ö n ? . 5 6
Adaptive Fuzzy Controllers (cont’d)
N
N
N
Σ
Σ
Inputlayer layer
fuzzificationnormalizationInference and
layer
Outputlayer
x2xnx1
�L(x)�2(x)�1(x) �y(1)1�y(L)1�y(1)m �y(L)m�y(2)m �y(2)1 y1ymWail Gueaieb 17 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Control SchemeLet A∗ denote the best possible approximation of matrix A in the face ofparametric uncertainties.
Then A’s modeling error can be expressed as A = A− A∗.
Hence, the ith manipulator control law may be reformulated as:
τi = D∗i (qi)qri
+ C∗i (qi, qi)qri
+ G∗i (qi)−Ksi
si − JTφi
(qi)(Kix + fdi)︸ ︷︷ ︸
τ(c)i
+ Di(qi)qri+ Ci(qi, qi)qri
+ Gi(qi)− τdi︸ ︷︷ ︸τ
(f)i
τ(c)i : CAC’s torque output (parametric uncertainties only)
τ(f)i : supervisory adaptive fuzzy regulator operating at a higher hierar-
chical level (lower bandwidth) than that of the CAC.
τ(f)i is modeled as an AFC à τ
(f)i = Ui(qi, qi, qri
, qri|Θi)
Wail Gueaieb 18 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Control Scheme (cont’d)
z AFC’s Computational complexity
τ(f)i = Ui(qi, qi, qri
, qri|Θi)
F 4 input vectors
F ki DOF for manipulator i
F κi membership functions to fuzzify each input element
à Total number of fuzzy rules fired by the AFC at manipulator i isLi = (κi)
4ki
For κi = 5 and ki = 3 à Li = 244, 140, 625 (too large!)
Wail Gueaieb 19 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Control Scheme (cont’d)
z Rule Decomposition Scheme
F Idea: aggregate Ui(qi, qi, qri, qri
|Θi) into several MIMO AFCs:
τ(f)i =
Ui(qi,qi,qri,qri
|Θi)︷ ︸︸ ︷Di(qi)qri
+ Ci(qi, qi)qri+ Gi(qi)− τdi︸ ︷︷ ︸
U2i (qi,qi,qri
)
F Unknown time-varying jth column of Di(qi) can be approximatedby a MIMO AFC U 1
ij(qi|Θ1ij).
à Di(qi)qri≈
∑ki
j=1 U 1ij(qi|Θ1
ij)qrij.
F qriis dependent on qi
à U 2i (qi, qi, qri
) can be replaced by U 2i (qi, qi|Θ2
i ).
F The torque offset generated by the supervisory adaptive fuzzyregulator module is:
τ(f)i =
ki∑j=1
U 1ij(qi|Θ1
ij)qrij+ U 2
i (qi, qi|Θ2i )
Wail Gueaieb 20 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Scheme (cont’d)
z Computational Complexity
F Each U 1ij(qi|Θ1
ij) fires L1ij = (κi)
ki rules.
F Number of rules fired by U 2i (qi, qi|Θ2
i ) is L2i = (κi)
2ki.
F Hence, the total number of rules fired by the AFC isLi = ki(κi)
ki + (κi)2ki for each robot.
+ Only [(κi)ki + (κi)
2ki] of them have different firing strengths(� (κi)
4ki).
F For κi = 5 and ki = 3 à number of distinct firing strengths tobe computed for each robot is 15, 750 (� 244, 140, 625).
Theorem 1 If the controller’s gains satisfy the required constraints, thenthe HIC gives rise to an asymptotic convergence of the payload’s positionand the internal forces tracking errors, x and fi, to zero.
Wail Gueaieb 21 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Scheme (cont’d)
z Numerical Results
F 5 Gaussian membership functions are used to fuzzify each input ofthe AFC module of the HIC.
F AFC module has no prior knowledge of the manipulators dynamics(i.e., Θ1
ij and Θ2i are initially set to zero for i = 1, 2 and j =
1, . . . , ki).
F 50% of the manipulators dynamics model is assumed to be known(for CAC.)
F For better computational efficiency, the supervisory adaptive fuzzyregulator module of the HIC is set to operate at a bandwidth 4times lower than that of the CAC.
Wail Gueaieb 22 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Scheme (cont’d)
KinematicsForward
High
Ban
dwid
th
CAC
Low
Expert Layer
Hard Algorithmic Layer
AFC
������
��� �����
�� �� �
�� �� �� ���� ��� � ���
�����
��� �
� �
� �
� � � �� � ��� �
� ����
��
�
� � �
� �!#"�
� � � �� � �$� ��� %'&)(*&)+', � ��� ��� -/.10�
-3240�
�5 "� � � ��� ����� � �6 "� � � � � �� ��������� � �7 "� � � ���
8:9 �;�<*= �> =� ; � � ��? @ =� ; � ����� A � �>CB� � � � � �� �D? @CB� �
Wail Gueaieb 23 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Scheme (cont’d)
y Experiment
F Intensity level of modeling uncertainties is varied by letting α spanthe interval [0, 2.5].
CAC’s Tracking Error HIC’s Tracking Error
05
1015
0
0.5
1
1.5
2
2.5−0.16
−0.14
−0.12
−0.1
−0.08
−0.06
−0.04
−0.02
0
0.02
Time (s)α
∆x (m
)
05
1015
0
0.5
1
1.5
2
2.5−0.15
−0.1
−0.05
0
0.05
Time (s)α
∆x (m
)
05
1015
0
0.5
1
1.5
2
2.5−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
Time (s)α
∆y (m
)
05
1015
0
0.5
1
1.5
2
2.5−0.12
−0.1
−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
Time (s)α
∆y (m
)
Wail Gueaieb 24 of 28
à áá à Ö n ? . 5 6
A Hybrid Adaptive Scheme (cont’d)
CAC’s Tracking Error HIC’s Tracking Error
05
1015
0
0.5
1
1.5
2
2.5−30
−20
−10
0
10
20
30
Time (s)α
∆ψ (d
eg)
05
1015
0
0.5
1
1.5
2
2.5−20
−15
−10
−5
0
5
10
15
Time (s)α
∆ψ (d
eg)
0 5 10 150
1
2
3−8
−6
−4
−2
0
2
4
6
8
Time (s)α
∆fi (N
)
0 5 10 150
1
2
3−8
−6
−4
−2
0
2
4
6
8
Time (s)α
∆fi (N
)
Wail Gueaieb 25 of 28
à áá à Ö n ? . 5 6
Conclusions
O Complex problem of controlling closed kinematic chain mechanisms:kinematic and dynamic coordination.
O Most adaptive controllers show success in the face of parametric un-certainties only.
O A novel hierarchical knowledge-based control scheme is proposed forthe control of CR.
O Innovative rule reduction technique is presented to significantly reducethe computational complexity.
O First attempt to control CR in the face of both structured and unstruc-tured uncertainties.
Wail Gueaieb 26 of 28
à áá à Ö n ? . 5 6
Conclusions (cont’d)
O Key characteristics of proposed hierarchical knowledge-based con-troller:
F Robustness in the face of parametric and modeling uncertainties ofvarying intensity levels.
F Both, position and internal force tracking errors are proven to con-verge to zero.
F Generic: easily portable from one platform to another (minor tun-ings may be needed.)
Wail Gueaieb 27 of 28
à áá à Ö n ? . 5 6
Future Research Directions
O Allow the automatic tuning of antecedent membership functions.
O Extend the zero-order Sugeno-type AFC to a first-order one: potentialof higher approximation capabilities.
O Extend the FLC model to a type-2 FLC to improve controllers robust-ness.
Wail Gueaieb 28 of 28