You are on page 1of 8

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/314673438

Computed Torque Control of the Stewart Platform with Uncertainty for Lower
Extremity Robotic Rehabilitation

Conference Paper · June 2017


DOI: 10.23919/ACC.2017.7963739

CITATION READS

1 230

4 authors:

Sahand Sab Arman Dabiri


The University of Arizona Southern Illinois University Edwardsville
10 PUBLICATIONS   23 CITATIONS    44 PUBLICATIONS   488 CITATIONS   

SEE PROFILE SEE PROFILE

David G. Armstrong Ali-Mohammad Poursina


Keck School of Medicine USC Universitetet i Agder
918 PUBLICATIONS   39,702 CITATIONS    51 PUBLICATIONS   317 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Surfactants and Biofilms View project

Amputation prevention In Harris Health System View project

All content following this page was uploaded by Arman Dabiri on 18 October 2018.

The user has requested enhancement of the downloaded file.


Computed Torque Control of the Stewart Platform with Uncertainty for
Lower Extremity Robotic Rehabilitation
Sahand Sabet1 , Arman Dabiri2 , David G. Armstrong 3 , and Mohammad Poursina 4

Abstract— Parallel manipulators play a key role in robotic from important properties such as high stiffness, high accu-
rehabilitation. In reality, such systems operate under uncer- racy, and large load-to-weight ratio. These qualities make it
tainty due to the changes in the characteristics of the patients a good candidate for being used as a rehabilitation robot.
and lack of knowledge about the physical and geometrical
properties of the system. In this paper, we present a robust Stewart platform was first used for rehabilitation purposes at
control scheme to control a six-degree-of-freedom Stewart Rutgers University [15]. This robot is also able to be operated
platform. In this application, it is aimed to follow a desired at home using telerahabilitation [2] which provides the ability
pure rotational motion required in the robotic rehabilitation of for remote monitoring and periodic reassessment. Several
the foot for patients with diabetic neuropathy. It is assumed that approaches can be used to perform the dynamic analysis and
uncertainty exists in the mass of the foot of the patients (the
proposed approach can also be used when disturbance exists). control [16] of the Stewart mechanism [17] including: virtual
To perform this, the method of polynomial chaos expansion work [18], Lagrange method [19], Newton Euler approach
(PCE) is extended and integrated with the computed torque [9], and Hamilton’s principle [11]. Therefore, the detail of
control law (CTCL) to control the system. In PCE scheme, deriving the dynamic equations of motion is not provided in
uncertainty is introduced to the system by compactly projecting this paper.
each stochastic response output and random input onto the
space of appropriate independent orthogonal polynomial basis In real applications, these systems operate under the in-
functions. CTCL uses a feedback linearization technique which fluence of uncertainty in the characteristics of the patients.
provides the necessary force/torque to enforce the system to In addition, lack of knowledge about the physical and
follow a prescribed trajectory. This papers presents a successful geometrical properties of the system is another source of
implementation of the PCE-base CTCL on a Stewart platform. uncertainty. Essentially, the designed platform is intended
Finally, a comparison between the efficiency and accuracy of
the Monte Carlo and PCE is conducted. to be used for the foot rehabilitation of the patients with
neuropathy. One of the most important uncertainties observed
I. I NTRODUCTION in this application is in the mass of the foot of the patient.
As such, the modeling scheme and controller should be
Multibody dynamics plays a key role in the modeling, robust enough to be able to model and control the platform
simulation, design, and control of many engineering prob- under uncertainty in the mass of the foot of the patients.
lems. One of the applications of multibody systems is in Considering uncertainty in the dynamic analysis and control
robotic rehabilitation. In this scheme, patients’ efforts can of multibody systems is relatively new. Polynomial Chaos
be evaluated, and therefore the training process can be Expansion (PCE) was first applied to the dynamic modeling
conducted to help them more effectively. Utilizing parallel of the nondeterministic systems by Sandu in [20], [21]. A
robots for rehabilitation purposes is relatively new [1], [2]. highly parallelizable divide-and-conquer based PCE has been
These robots are capable of doing repeated tasks with high developed in [22] to reduce the computational complexity
precision. Valuable efforts and advances have been conducted of the simulation of nondeterministic multibody systems
to the development of the parallel mechanisms [3]–[11]. [23]. In the field of control of multibody systems, PCE
Among parallel mechanisms, the Stewart platform is the most was first applied to control nondeterministic systems in
celebrated one which was introduced in 1965 [12]. Major 2003 [24]. Templeton has applied the same technique to an
industrial applications of this device are flight simulators, au- H2 optimal control problem [25]. Voglewede et. al., have
tomobile simulators, alignment mechanisms. In our case, the applied polynomial chaos theory to control a SCARA robot
Stewart platform is used as a neurorehabilitative exergaming manipulator with uncertainties in the mass of the constituent
platform [13], [14]. This complex multibody system benefits bodies [26]. A general scheme to apply PCE based CTCL
to nondeterministic multibody systems has provided in [23].
1 Sahand Sabet is with the Aerospace and Mechanical
Engineering Department, University of Arizona, AZ 85721, USA
This paper considers the control of this platform with
sahandsabet@email.arizona.edu uncertainty in the mass of the foot of the patient. As such,
2 Arman Dabiri is with the Aerospace and Mechanical Engineering
the Stewart platform with the 6-UPS mechanical structure
Department, University of Arizona, AZ 85721, USA (corresponding author: as shown in Fig. 1 is used for this analysis. It consists of a
Phone: 520-561-5612; armandabiri@email.arizona.edu)
3 Prof. Armstrong is with the Department of Surgery, top platform, a base platform, and six linear actuators (legs).
College of Medicine, University of Arizona, AZ 85721, USA Each actuator is connected to the base plate and the top
dga@email.arizona.edu plate with a universal joint and a spherical joint, respectively.
4 Prof. Poursina is with the Aerospace and Mechanical
Engineering Department, University of Arizona, AZ 85721, USA The radius of the top plate, the radius of the platform base,
mpoursina@gmail.com the height of the system, and the length of the links are
Nt
R(t, ζ) ∼
X
= ri (t)Λi (ζ). (2)
i=0
For a case with Nu number of uncertain parameters ζ =
(ζ1 , ζ2 , ..., ζNu ), the total number of terms Nt + 1 in (2)
with maximum order of polynomial P can be evaluated as
[20], [22]

(Nu + P )!
Nt + 1 = . (3)
Nu !P !
The orthogonal base functions presented in (1) can
be chosen based on the type of random variable and the
support range as can be seen in Tab. I. According to [29],
these polynomials establish a complete orthogonal basis for
the Hilbert space of square integrable random variables.
Therefore, the weighted inner product of two base functions
Fig. 1: A Stewart platform with the 6-UPS parallel manipulators. for one-dimensional random variable can be evaluated as
Z
< λi (ζ), λj (ζ) > , λi (ζ)λj (ζ)w(ζ)dζ = h2i δij , (4)
denoted by r, R, h, and li , i = 1, 2, · · · , 6, respectively. The
offset angles θt and θb on the both top and base plates are where δij is the Kronecker delta and h2i is a positive number.
also illustrated in the figure. Herein, the method of PCE is When considering n-dimensional random variable, (4) can
extended and integrated with the Computed Torque Control be represented as
Law (CTCL). In PCE scheme, uncertainty is introduced to < Λ (ζ), Λj (ζ) >=
the system by compactly projecting each stochastic response Z iZ
output and random input onto the space of appropriate ... Λi (ζ)Λj (ζ)w(ζ1 )...w(ζn )dζ1 ...dζn = c2i δij , (5)
independent orthogonal polynomial basis functions. CTCL
where c2i is a positive number.
presents a feedback linearization technique which provides
Eventually, having the PCE-series coefficients ri corre-
the necessary force/torque to enforce the system to follow
sponding to the orthogonal base functions Λi (ζ1 , ..., ζn ), one
a prescribed trajectory. A general scheme is introduced to
can easily evaluate the stochastic response of the system
solve the stochastic equations of motion. At the end, this
including expected value and standard deviation using the
method is applied to control a Stewart mechanism with the
following equations
uncertainty described previously. Herein, a PD controller is
used. However, using the proposed method, a PID controller E[R] = r0 , (6)
could be utilized when disturbance exists. Finally, the results ∞
X
of the comparison between the traditionally used Monte V ar(R) = E[R2 ] − (E[R])2 = ri2 < Λ2i > . (7)
Carlo method and the intrusive PCE in the context of the i=1
time efficiency and accuracy is presented. A detail explanation to find the PCE-series expansion
coefficients ri will be provided in this paper.
II. P OLYNOMIAL C HAOS E XPANSION FOR U NCERTAINTY
A NALYSIS III. C OMPUTED T ORQUE C ONTROL L AW FOR
PCE provides a convenient framework to compactly D ETERMINISTIC S YSTEMS
project each stochastic response output and random input This section explains the Computed Torque Control Law to
onto the space of orthogonal base functions as [20], [27] control deterministic multibody systems. CTCL is a feedback

X linearization method which uses inverse dynamic to find
R(t, ζ) = ri (t)Λi (ζ). (1) the required force to make the system follow a prescribed
i=0 trajectory [30]. The dynamic equations of a fully-actuated
In this equation, R(t, ζ) is a random quantity, ri rep- deterministic multibody system can be stated as
resents the time-dependent coefficients, and Λi (ζ1 , ..., ζn )
M (q)q̈ + N (q, q̇) + fd = f, (8)
(usually denoted as λi (ζ) in case of one-dimensional random
variable) contain the time-invariant generalized polynomial where M (q) is the inertia matrix of the entire system, q
chaos. These multidimensional polynomials are expressed in denotes the column matrix of the joint variables, and f
terms of random variables ζ(ζ1 , ..., ζn ). The series shown in represents column matrix of the generalized driving/control
(1) converges to any random process in L2 sense [28]. One force [31] at each joint. The term N (q, q̇) includes the non-
might truncate this infinite expansion at a finite number of control generalized forces due to Coriolis/centripetal and
terms Nt as gravity terms [32] and fd denotes the disturbance vector
TABLE I: RELATION BETWEEN BASE FUNCTIONS, SUPPORTS, AND DISTRIBUTION FUNCTIONS

Distribution Density function Base function Support range Weight function


1
Uniform 2 Legendre [−1, 1] 1
(1−x) (1+x)β
α
Beta 2( α+β+1)B(α+1,β+1)
Jacobi [−1, 1] (1 − x) (1 + x)β
α

−x2 −x2
Normal √1 e 2 Hermite (−∞, ∞) e 2

−x
Exponential e Laguerre [0, ∞) e−x
xα e−x
Gamma Γ(α+1) Generalized Laguerre [0, ∞) xα e−x

which is assumed to be zero. Defining the prescribed joint input w can be substituted into (16) to compute the driving
trajectory as qd (t), one can express the tracking error at force f as
position, velocity and acceleration levels as [23]
f = M (q)(q̈d + Kv ė + Kp e) + N (q, q̇). (20)
e(t) = qd (t) − q(t), (9)
The driving force f in (20) can be substituted into (17),
ė(t) = q̇d (t) − q̇(t), (10) resulting in the closed-loop dynamic equation as
ë(t) = q̈d (t) − q̈(t). (11)
M (q)q̈ + N (q, q̇) = M (q)(q̈d + Kv ė + Kp e) + N (q, q̇), (21)
The dynamic equation of system (8) can be solved for q̈ and
the result can be substituted into (11), resulting in which can be simplified as

ë(t) = q̈d (t) + M −1 (q)(N (q, q̇) − f ). (12)


ë + Kv ė + Kp e = 0. (22)
By defining the control input function as
It can be observed that the highly nonlinear dynamic equation
w = q̈d (t) + M −1 (q)(N (q, q̇) − f ), (13) of system ( (8)) is transformed to a linear ODE system. The
closed-loop dynamic equation of system can be presented
One can define the state x as
in state space transformation as [32] where Kp and Kv are

e(t)
 diagonal gain matrices [23].
x= . (14)
ė(t)     
The time derivative of x can be expressed in Brunovsky ė(t) 0 I e(t)
= . (23)
canonical form [32], by using (10 - 13), as ë(t) −Kp −Kv ė(t)

ė(t)
 
0

I e(t)
  
0 Where this system is asymptotically stable as long as all
= + w. (15) of the gain coefficients are positive [32].
ë(t) 0 0 ė(t) I
This closed-loop control scheme shown in Fig. (2) is com-
The above equation can be stabilized and the tracking error prised of two components [33]. The first component consists
e is driven to zero by choosing an appropriate input function of a nonlinear inner loop where the generalized applied
w. Equation (13) can then be solved to find the generalized forces (actuator torques) are calculated and the equations of
driving forces f as motion are solved by an integrator. This part of the controller
design is essentially the inverse dynamics of the system.
f = M (q)(q̈d − w) + N (q, q̇). (16)
The second component is an outer loop which stabilizes the
As such, by using a control input in (15), the nonlinear system by using PI, PID, or any other type of controller that
control input given by f drives the multibody system to makes the system follow the specified trajectory [30]. The
follow the prescribed trajectory defined at the joints of the outer-loop signal w can be chosen using many approaches,
system [32]. This is called Computed Torque Control Law. including robust and adaptive control techniques [32].
Substituting (16) into (8) one can represent the dynamic
equation of the closed-loop control system as IV. C OMPUTED T ORQUE C ONTROL L AW FOR
N ONDETERMINISTIC S YSTEMS
M (q)q̈ + N (q, q̇) = M (q)(q̈d − w) + N (q, q̇), (17)
This section provides the framework to apply the PCE
⇒ ë = w. (18)
method to the Computed Torque Control Law. The dynamic
In this paper, a PD controller in the following format is equation of a nondeterministic multibody system can be
used to control the system (a PID controller could be used stated as
if uncertainty existed) ˙ = f˜,
M̃ (q̃)q̃¨ + Ñ (q̃, q̃) (24)
w = −Kv ė − Kp e, (19)
where the tilde sign shows that these terms contain stochastic
where Kv and Kp are diagonal gain matrices. The control values. By using the same defined term in (9 - 15 ), one might
and makes it independent of uncertainty in the geometry [23].
Inverse Linear
System A. Computing Nondeterministic Generalized Driving Forces
Dynamics 𝑁(𝑞, 𝑞)
𝑞𝑑 + + 𝑞, 𝑞
The nondeterministic driving force f˜ at any time can be
𝑓 Forward
𝑀(𝑞) + Dynamics found by using (25). The results can then be shown in the
- +
PCE scheme as
𝑞, 𝑞 fd
Nt
X
𝑤 Outer Loop f˜ = [f˜1 , ..., f˜n ]T = F̃m Λm (ζ), (30)
Feedback m=0
𝑞𝑑, 𝑞𝑑
where n represents the size of the force vector f˜, the term
Nt is the total number of terms in the summation, and
Fig. 2: Different components of computed torque control method F̃m = [f˜1m , ...., f˜nm ]T , m = 0, ..., Nt . (31)
to control multibody systems [32]
The expansion coefficients associated with each control force
in (31) can be calculated using the Galerkin projection
technique. As an example, suppose that at a given time, the
find the generalized driving force f˜ as j th generalized driving force (f˜j ) is evaluated as a function
f˜ = M̃ (q)(q̃¨d − w̃) + Ñ (q̃, q̃).
˙ (25) of ζ as gj (ζ). The generalized driving force can be displayed
in terms of its expansion coefficients as [23]
Using a PD controller as an input function w̃
Nt
X
w̃ = −Kv ẽ˙ − Kp ẽ (26) f˜j = f˜jm Λm (ζ) = gj (ζ). (32)
m=0
One may express the generalized driving force stated at (25) The projection of the above relation onto the space of a
as designated base function Λl is expressed as
f˜ = M̃ (q̃)(q̃¨d + Kv ẽ˙ + Kp ẽ) + Ñ (q̃, q̃).
˙ (27)
< f˜j , Λl (ζ) > = < gj , Λl (ζ) > . (33)
The tracking error dynamics is then revised as
On the other hand,
ẽ¨ = q̃¨d (t) − q̃(t) ˙
¨ = −Kp ẽ − Kv ẽ, (28)
Nt
X
This equation shows that the existence of uncertainty in < f˜j , Λl (ζ) > = < f˜jm Λm (ζ), Λl (ζ) >
the geometry of the system does not affect the closed-loop m=0
Nt
equation. X
For the case with no uncertainty in the initial joint variable = c2l f˜jm δml
m=0
qt0 , its time derivative q̇t0 , and the desirable joint motion qd ,
the closed-loop control equation can be simplified as = c2l f˜jl , (34)
where l = 0, ..., Nt , and j = 1, ..., n, while c2l is defined as
ë + Kv ė + Kp e = 0. (29)
c2l = < Λl (ζ), Λl (ζ) > l = 0, ..., Nt (35)
As can be seen, the above equation does not include
any stochastic term. One important key point that can be Therefore, the PCE-series expansion coefficients for f˜j can
observed is that the closed-loop control equation in the be calculated as
nondeterministic case (29) with deterministic qt0 and q̇t0 1
and nondeterministic physical and geometrical properties of f˜jl = < gj , Λl (ζ) >,
c2l
the bodies of the system is exactly same as the deterministic l = 0, ..., Nt , and j = 1, ..., n. (36)
case ( (22)). In this situation, the joint variable q and its time
derivative q̇ can be evaluated as a function of time by solving Having the PCE-series expansion coefficients f˜jl , the
the deterministic ODE stated in (29). Having q and q̇, one mean and standard deviation of generalized forces can be
can find the matrices M̃ and Ñ at any given time. Then, the easily calculated using (6) and (7).
generalized driving force f˜ can be evaluated at any time step V. R ESULTS
by using (25).
It needs to be mentioned that even though in this situation, A. System Description
the closed-loop control equations are the same for both This section considers the application of CTCL in con-
deterministic and nondeterministic cases, the generalized trolling a Stewart platform shown in Fig. 1 with uncertainty
driving force are different. For the nondeterministic case in the mass of the foot. It is assumed that the foot is rigidly
(25) the driving forces f˜ contain stochastic terms (due to attached to the center of mass of the top plate of the platform,
the existence of uncertainty in M̃ and Ñ matrices). In other and has the same rotation and translation as the top plate.
words, applying the generalized driving force f˜ at each time The generalized coordinates for this problem are defined as
step linearizes the closed-loop control equation of the system, the lengths of the legs (shown in Fig. 1). The top plate’s
TABLE II: PHYSICAL AND GEOMETRICAL PROPERTIES OF
THE STEWART PLATFORM
of motion of the ankle is conservatively equal to:
π π
Parameter Value Distribution of Uncertainty θxd = 30 sin( t) (rad), (37)
180 2
ρtop platf orm 2700(Kg/m3 ) N one π π
m̃top platf orm 25 + 5ζ1 (Kg) ζ1 ∈ [−1, 1]
θyd = 30 sin( t) (rad),
180 2
R 1(m) N one π π
r 0.325(m) N one
θzd = 30 sin( t) (rad).
180 2
θtop platf orm 0.225(Degree) N one This trajectory is required for the rehabilitation where the
θbottom platf orm 0.5(Degree) N one range of motion for the ankle is founded in [34]. It is assumed
H 0.5(m) N one
that uncertainty with uniform distribution exists in the mass
of the foot. Therefore, Legendre polynomials are used as the
appropriate base functions for this problem (Tab. I).
B. Evaluating Joint Coordinates and Speeds
TABLE III: INITIAL CONDITIONS AND DESIRABLE MOTION
OF THE TOP PLATFORM CENTER OF MASS In this paper, it is assumed that uncertainty only exists in
the mass of the platform. Therefore, the closed-loop control
Parameter Value equation of the system, (29), is not affected by uncertainties.
x0 0.01(m) Accordingly, (29) is used for both Monte Carlo and PCE
y0 0.01(m) approaches. This equation is solved for e and ė using ODE45
z0 0.1(m) for the time period of t ∈ [0, 2] (sec). Eventually, the gener-
θx0 10(Degree) alized coordinate of the system, q, and their time derivative,
θ y0 10(Degree) q̇, are evaluated using (9 - 10). Knowing the values of the
θz0 10(Degree) generalized coordinates, the position and orientation of the
ẋ0 0(m/s) center of mass of the upper plate are evaluated, and shown
ẏ0 0(m/s) in Fig. 3 (the plots are not shown for all of the variables in
ż0 0(m/s) order to save space). It is observed that after approximately
θ̇x0 5π(Degree/Sec) 1.3 second all of the joints follow the prescribed motion.
θ̇y0 5π(Degree/Sec) C. Evaluating the Nondeterministic Generalized Driving
θ̇z0 5π(Degree/Sec) Forces Using PCE and Monte Carlo Method
xd 0(m)
Having the generalized coordinate q (L1 , L2 , ..., L6 ) and
yd 0(m)
its time derivative q̇ (L̇1 , L̇2 , ..., L̇6 ), the required actuator
zd 0(m)
forces to make the system follow the desirable paths can be
θxd 30sin(tπ/4)(Degree)
calculated. A traditionally used Monte Carlo method ( (20))
θ yd 30sin(tπ/4)(Degree)
with 104 runs is used to find the driving forces f by sweeping
θzd 30sin(tπ/4)(Degree)
the uncertainty in the mass of the foot. Then, the mean value
and standard deviation of the generalized driving forces are
shown in Figs. 4 and 5. In addition, (27) is used in order to
evaluate the generalized forces in the PCE scheme.
Since the stochastic mass of the top plate (which includes
physical and geometrical properties are provided in Tab. II. the mass of the foot as well) in this problem is stated in the
The desirable motion for the center of mass of the upper polynomial forms (mtop = 25 + 5ζ1 ), the elements of the
plate of the platform (shown with subscript d) and its initial matrices M̃ (q) and Ñ (q, q̇) will have polynomial form as
conditions are provided in Tab. III. Six PD controller with well. It was observed that the maximum order of polynomial
the following gain values are chosen to control each leg: in the elements of matrices M̃ (q) and Ñ (q, q̇) is equal to 1.
Kv = 10I(6×6) , Kp = 25I(6×6) . Considering uncertainty in Eventually these polynomial terms in M̃ (q) and Ñ (q, q̇) are
the mass of the upper plate, it is desired to find the driving multiplied by deterministic values (q̈d +Kv ė+Kp e) in (27) to
forces in the actuator (mean value and standard deviation) at make the driving forces f˜. Consequently, the maximum order
any given time due to the stochastic behavior of the system’s of the driving forces f˜ will remain 1. Therefore, the one-
dynamics. dimensional Legendre polynomial of order 1 with 2 terms
(1, ζ1 ) can capture all stochastic elements of the driving
In this research, the mass M̃ and Coriolis/centripetal force f˜ by projecting it into the space of orthogonal base
matrix Ñ are explicitly evaluated as functions of translational functions using (30 - 36). Therefore, using the PCE method,
and rotational coordinates of the center of mass of the top one can find the exact order of polynomial needed and there
plate using Lagrange method. The elements of these matrices will be no need to conduct a trial and error. Eventually, the
are not shown in this paper due to the existence of excessive mean value and standard deviation of driving force f˜ can be
elongate terms. It is assumed that the translational coordinate evaluated using (6 - 7). It needs to be mentioned that using
of the top plate remains stationary and the rotational range the Monte Carlo method, in order to evaluate the mean value
-40

Mean Value of Generalized Forces f~ (N)


-60

0.5
-80
Trajectory of the Center of Mass

-100
0.4

-120
Trajectory of the -rst joint x1 ( from PCE and Monte Carlo)
0.3 Desirable trajectory of the -rst joint x1 = 0 -140
Trajectory of the -rst joint 31 ( from PCE and Monte Carlo)
Mean Value of f~1 from Monte Carlo
Desirable trajectory of the -rst joint 31 = 30:/180sin(2t:/4)
-160 Mean Valueof f~1 from PCE
0.2 Mean Valueof f~6 from Monte Carlo
-180 Mean Valueof f~6 from PCE

0.1 -200

-220
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0 Time (sec)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (sec)

Fig. 4: Mean value of the generalized forces

Fig. 3: JOINTS’ TRAJECTORIES


VI. C ONCLUSION
In this paper, the intrusive polynomial chaos expansion
and standard deviation of the driving forces f at each time has been applied to the Computed Torque Control of non-
step, (27) need to be evaluated for 104 times. While in the deterministic Stewart platform. Since this parallel robot is
PCE scheme, the driving force needs to be evaluated only used for the robotic rehabilitation, uncertainty in the mass
once at each time step. This is the key fact that makes the of the foot of the patients have been considered in the
PCE faster than the Monte Carlo approach. modeling and control. To perform this robust control scheme,
it has been assumed that the foot can be modeled as a mass
D. Accuracy and Efficiency rigidly attached to the top plate of the platform. The uniform
distribution in the mass of the foot has been considered
To compare the accuracy of the PCE and the Monte for this analysis. It has been demonstrated that how PCE
Carlo method, the mean value and standard deviation of can compactly model uncertainty in the system. Further,
generalized driving forces are shown in Figs. 4-5. Since the mathematical formulations have been generated to formulate
PCE results provide the exact solution, the relative error is CTCL in the PCE format. The presented PCE-based CTCL
evaluated by comparing the Monte Carlo results with the has been successfully used to control the Stewart platform
PCE ones as with the uncertainty in the mass of the foot to follow a
Percentage of Relative Error = ||x1 − y1 || × 100, (38) desired trajectory. At the end, a comparison of the time
efficiency and accuracy between the traditionally used Monte
where x1 and y1 are respectively the obtained results from Carlo method and the intrusive PCE has been presented. The
the 1.0 × 104 Monte Carlo simulations and the PCE of order evaluated results indicate that the intrusive PCE approach can
1. The mean value of the generalized force f˜ from both provide better accuracy with much less computation time
PCE and Monte Carlo is shown in Figs. 4. Both methods than the Monte Carlo method.
provide identical values in the mean value (error up to 10−9 ).
Fig. 5 shows the standard deviation of the generalized forces VII. ACKNOWLEDGMENT
f˜ obtained from both PCE and Monte Carlo methods. The The authors would like to acknowledge the University of
Monte Carlo is able to provide the standard deviation of Arizona for the Accelerate for Success Grant to perform this
the driving forces with a relative error (using (38)) up to project.
0.43 percents. To evaluate the efficiency of the methods,
R EFERENCES
the computation time to solve the dynamic equation of
system as well as evaluating the stochastic response of the [1] H. I. Krebs, B. T. Volpe, M. L. Aisen, and N. Hogan, “Increasing
productivity and quality of care: Robot-aided neuro-rehabilitation,”
driving force f˜ is evaluated using both PCE and Monte Journal of Rehabilitation Research and Development, vol. 37, pp. 639–
Carlo methods. It was observed that PCE has a computational 652, 2000.
time of 64.08 second, while Monte Carlo run time is 95.30 [2] M. Girone, G. Burdea, M. Bouzit, V. Popescu, and J. E. Deutsch,
“Stewart platform-based system for ankle telerehabilitation,” Au-
second. Therefore, PCE is able to provide a better accuracy tonomous Robots, vol. 10, pp. 203–212, 2001.
with shorter computation time. It needs to be mentioned [3] K. Hunt, “Contribution to discussion of papers on research in auto-
that the difference between the computation time of these mobile stability, control and tyre performance,” in Proceedings of the
the Automobile Division institution, 1956-1957.
methods will increase as the number of uncertain parameter [4] R. Clavel, “Delta, a fast robot with parallel geometry,” 18th Interna-
Nu increases (a detail explanation can be found in [23]). tional Symposium on Industrial Robot, no. 91100, pp. 206–210, 1988.
Standard Deviation of Generalized Forces f~ (N) [20] A. Sandu, C. Sandu, and M. Ahmadian, “Modeling multibody systems
25
with uncertainties. part I: Theoretical and computational aspects,”
Multibody System Dynamics, vol. 15, no. 4, pp. 369–391, 2006.
[21] C. Sandu, A. Sandu, and M. Ahmadian, “Modeling multibody systems
20 with uncertainties. part ii: Numerical applications,” Multibody System
Dynamics, vol. 15, no. 3, pp. 241–262, 2006.
[22] M. Poursina, “Extended divide-and-conquer algorithm for uncertainty
15
analysis of multibody systems in polynomial chaos expansion frame-
work,” Journal of Computational and Nonlinear Dynamics, vol. 11, 5
Standard Deviation of f~1 from Monte Carlo 2016.
Standard Deviation of f~1 from PCE [23] S. Sabet and M. Poursina, “Robust framework for the computed
10
Standard Deviation of f~6 from Monte Carlo torque control of nondeterministic multibody dynamic systems,” in
Standard Deviation of f~6 from PCE Proceedings of the ASME 2015 International Design Engineering
Technical Conferences and Computers and Information in Engineering
5 Conference, (Charlotte, North Carolina), August 2016.
[24] Z. K. Nagy and R. D. Braatz, “Recent advances in the optimal control
of batch processes,” Recent Research Developments in Chemical
0
Engineering, vol. 5, no. 1, pp. 99 – 127, 2003.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
[25] B. A. Templeton, “”a polynomial chaos approach to control design”,
Time (sec)
PhD Thesis, Virginia Polytechnic Institute and State University, VA,
August,” 2009.
[26] P. Voglewede, A. smith, and A. Monti, “Dynamic performance of
Fig. 5: Standard deviation of the generalized forces a scara robot manipulator with uncertainty using polynomial chaos
theory,” IEEE Transaction on Robotics, vol. 25, no. 1, pp. 206–210,
2009.
[27] M. Poursina, “Efficient application of polynomial chaos expansion
[5] F. Pierrot, A. Fournie, and P. Dauchez, “Hexa: a fast six-dof fully- for the dynamic analysis of multibody systems with uncertainty,,” in
parallel robot,” Fifth International Conference on Advanced Robotics Proceeding of ASME 2014 International Design Engineering Technical
Robots in Unstructured Environments,, vol. 2, no. 1, pp. 1158 – 1163, Conferences, Computers and Information in Engineering Conference,
1991. Buffalo, NY, (Buffalo, NY), 2014.
[6] G. Lebret, K. Liu, and F. L. Lewis, “Dynamic analysis and control of [28] R. H. Cameron and W. T. Martin, “The orthogonal development of
a stewart platform manipulato,” Journal of Robotic systems, vol. 10, non-linear functionals in series of fourier-hermite functionals,” The
no. 5, pp. 711 – 721, 1992. Annals of Mathematics, vol. 48, no. 2, pp. 385–392, 1947.
[7] H. Abdellatif and B. Heimann, “Computational ecient inverse dy- [29] D. Xiu and G. E. Karniadakis, “The wiener–askey polynomial chaos
namics of 6-dof fully parallel manipulators by using the lagrangian for stochastic differential equations,” SIAM Journal of Scientific Com-
formalism,” Mechanism and Machine Theory, vol. 44, p. 192207, puting, vol. 24, no. 2, pp. 619–644, 2002.
2009. [30] C. Kingsley, “Efficient computation of inverse dynamics for computed
[8] Y. Nakamura and M. Ghodoussi, “Dynamics computation of closed- torque control applications of fully actuated complex multibody sys-
link robot mechanisms with nonredundant and redundant actuators,” tems,” Master’s thesis, The University of Arizona, Tucson, Az, 2015.
IEEE Transactions on Industrial Electronics, vol. 5, no. 3, p. 294302, [31] P. Masarati, “Computed torque control of redundant manipulators
1989. using general-purpose software in real-time,” Multibody System Dy-
[9] B. Dasgupta and P. Choudhury, “A general strategy based on the namics, vol. 32, no. 4, pp. 403–428, 2014.
newton-euler approach for the dynamic formulation of parallel manip- [32] F. L. Lewis, D. M. Dawson, and C. T. Abdallah, Robot Manipulator
ulators,” Mechanism and Machine Theory, vol. 34, no. 6, pp. 801–824, Control: Theory and Practice. Marcel Dekker, Inc, 2003.
1993. [33] S. Sabet, “Kinematic analysis and inverse dynamics-based control of
[10] R. Clavel, Conception dun robot parall‘ele rapide ‘a4degres de liberte. nondeterministic multibody systems,” Master’s thesis, The University
PhD thesis, EPFL, 1991. of Arizona, Tucson, AZ, 2016.
[11] K. Miller, “Optimal design and modeling of spatial parallel manip- [34] S. Siegler, J. Chen, and C. Schneck, “The three-dimensional kinemat-
ulators,” International Journal of Robotics Research, vol. 23, no. 2, ics and flexibility characteristics of the human ankle and subtalar joints
pp. 127–140, 2004. part i: Kinematics,” Journal of Biomechanical Engineering, vol. 110,
[12] D. Stewart, “A platform with 6 degrees of freedom,” pp. 371 – 386, no. 4, pp. 364–373, 1988.
1965.
[13] G. S. Grewal, M. Schwenk, J. Lee-Eng, S. Parvaneh, M. Bharara, R. A.
Menzies, T. K. Talal, D. G. Armstrong, and B. Najafi, “Sensor-based
interactive balance training with visual joint movement feedback for
improving postural stability in diabetics with peripheral neuropathy: a
randomized controlled trial,” Gerontology, vol. 61, no. 6, pp. 567–574,
2015.
[14] G. S. Grewal, R. Sayeed, M. Schwenk, M. Bharara, R. Menzies,
T. K. Talal, D. G. Armstrong, and B. Najafi, “Balance rehabilitation:
promoting the role of virtual reality in patients with diabetic peripheral
neuropathy,” Journal of the American Podiatric Medical Association,
vol. 103, no. 6, pp. 498–507, 2013.
[15] M. J. Girone, G. C. Burdea, and M. Bouzit, “’rutgers ankle’ orthopedic
rehabilitation interface,” vol. 67, pp. 305–312, 1999.
[16] R. Gexue, L. Qiuhai, H. Ning, N. Rendong, and P. Bo, “On vibration
control with stewart parallel mechanism,” Mechatronics, vol. 14, no. 1,
pp. 1 – 13, 2004.
[17] J.-P. Merlet, Parallel Robots. Springer, 2006.
[18] L. V. Tsai, “Solving the inverse dynamics of a stewart-gough manip-
ulator by the principle of virtual work,” ASME Journal of Mechanical
Design, vol. 122, no. 1, p. 39, 2000.
[19] K. Liu, G. Lebret, and D. Taylor, “The singularities and dynamics
of a stewart platform manipulator,” Journal of Intelligent and Robotic
Systems, vol. 8, p. 287308, 1993.

View publication stats

You might also like