You are on page 1of 5

206

IEEE TRANSACTIONS ON ROBOTICS, VOL. 25, NO. 1, FEBRUARY 2009

Dynamic Performance of a SCARA Robot Manipulator With Uncertainty Using Polynomial Chaos Theory
Philip Voglewede, Anton H. C. Smith, and Antonello Monti
AbstractThis short paper outlines how polynomial chaos theory (PCT) can be utilized for manipulator dynamic analysis and controller design in a 4-DOF selective compliance assembly robot-arm-type manipulator with variation in both the link masses and payload. It includes a simple linear control algorithm into the formulation to show the capability of the PCT framework. Index TermsMonte Carlo methods, polynomials, robot dynamics, uncertainty.

be utilized to better understand and predict dynamic performance of manipulators under feedback control with uncertainty. II. PCT BASICS The basis of PCT is the expansion all the generic variables in the governing equations in terms of a orthogonal polynomial basis. Following a notation similar to [10], a generic variable u can be expressed as a innite sum of orthogonal polynomials as

u ( ) = u 0 I0 +
i1 = 1

ui 1 I1 (i 1 ( ))
i1

+ I. INTRODUCTION The design of any mass-produced system involves understanding how manufacturing variation will affect its performance. For robotic manipulators, this problem is exacerbated by changes in the payload. While these affects are minimal for highly geared manipulators, the affects are large for direct driven manipulators. Robotic designers need to better understand how variation in both product manufacture and payload affects the dynamic performance of these types of manipulators. Unfortunately, there is a lack of appropriate tools to generally assess the dynamic performance of mechanisms with uncertainty. There has been signicant work on the study of how clearances statically affect the output of a mechanism [1][4]. There are also numerous computer-aided tolerancing tools that can perform static stack-ups via linearization or a statistical Monte Carlo (MC) analysis [5][8]. However, there is a lack of appropriate tools to simulate and predict the window of possible dynamic congurations resulting from the variation. As such, variational dynamic simulations are typically performed as an MC of the governing equations that, in turn, results in large run times. Polynomial chaos theory (PCT) can be used for uncertain dynamic analysis, including the controller design. PCT allows one to solve stochastic differential equations by using orthogonal polynomials in conjunction with a Galerkin projection. This allows the transformation of stochastic differential equations into an expanded set of standard ordinary differential equations [9][12]. PCT has been successfully applied to uid mechanic [13], general oscillatory [14], heat transfer [15], and electrical [16] systems. More recently, it has been applied specifically to multibody dynamic systems [17][19], but specic examples outlining uses in this realm are lacking. This paper applies PCT to a selective compliance assembly robot arm (SCARA)-type manipulator to show how the methodology can
i1 = 1 i2

ui 1 i 2 I2 (i 1 ( ), i 2 ( )) +

(1)

where is the random event, = {1 , 2 , . . . , n } is the random vector, Ii is the orthogonal polynomial in the Askey scheme, and ui , ui j , etc., are denoted as the PC coefcients. Since this notation is rather cumbersome to deal with, it is usually abbreviated as

u ( ) =
j=0

uj j ( ( ))

(2)

where uj are the PC coefcients and j ( ) are the multidimensional orthogonal polynomials. For practical purposes the polynomial expansion is truncated to a nite number of polynomial terms and the dependence on the uncertain event is assumed. How many terms are kept is dependent upon the number of uncertainties in the system nv and the order of the adopted polynomial np .1 The total number of terms of the expansion P is given by P = (np + nv )! n p !n v ! 1. (3)

The expansion of the variables leaves the unknown vector that must be dealt with. However, due to the special nature of the orthogonal polynomials, the Galerkin projection of the expanded equations will eliminate the dependence upon the unknown stochastic variable . This is done by dening the inner product as f ( )g ( ) = f ( )g ( )W ( ) d (4)

where W ( ) is the weighting function associated with the chosen basis in the Askey scheme (see [10] for detailed descriptions). Due to the special nature of the inner product, many of the inner products will be zero as (5) i j = 2 i i j where i j is the Kronecker delta. The results are an expanded set of equations that can be solved using traditional techniques for the unknown PC coefcients. The PC coefcients can then be backsubstituted into (2) to yield the distribution of the original variables in time. A. PCT in Manipulator Dynamic Analysis Unfortunately, some nonlinearities cannot be solved using PCT, as the Galerkin projection does not eliminate the dependence upon the unknown variables i . For multidimensional serial manipulators, the general form of the dynamic equations of motion follow [20] + C (, ) + G ( ) = M
1 See

Manuscript received May 25, 2007; revised September 2, 2008. First published January 13, 2009; current version published February 4, 2009. This paper was recommended for publication by Associate Editor I. Chen and Editor F. Park upon evaluation of the reviewers comments. This work was supported by the U.S. Ofce of Naval Research under Grant N00014-02-1-0623. P. Voglewede is with the Mechanical Engineering Department, Marquette University, Milwaukee, WI 53132 USA (e-mail: philip.voglewede@marquette. edu). A. H. C. Smith is with the Electrical Engineering Department, University of South Carolina, Columbia, SC 29208 USA (e-mail: smith35@cec.sc.edu). A. Monti is with the Institute for Automation of Complex Power Systems at the EON Energy Research Center, Rhenish-Westphalian Technical University (RWTH) Aachen, Aachen, Germany (e-mail: amonti@eonerc.rwth-aachen.de). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TRO.2008.2006871

(6)

[10] for a discussion on the convergence of this series.

1552-3098/$25.00 2009 IEEE


Authorized licensed use limited to: RWTH AACHEN. Downloaded on February 18, 2009 at 03:05 from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON ROBOTICS, VOL. 25, NO. 1, FEBRUARY 2009

207

TABLE I SCARA PARAMETERS FOR NUMERICAL STUDY

Fig. 1. Schematic of the SCARA robot showing the geometric parameters and coordinate frames.

in Fig. 1 are found to be [21], [22] p 3 + 0 .5 p 2 c 2 0 p5 1 p1 + p2 c2 p 3 + 0 .5 p 2 c 2 p3 0 p5


M

0 0 p4 0

where M is the mass matrix, C is a vector of Coriolis and centrifugal terms, G is the gravitational terms, and is a vector of generalized forces. These terms, in general, contain polynomial and trigonometric nonlinearities. Polynomial nonlinearities cause no problems to the formulation. Trigonometric functions can be solved easily using a Taylor series expansion with a small-angle approximation, as will be shown later in this paper.

1 p5 2 0 3 p5 4 p5 0 0 0 0 0 1 0 0 0 2 + 0 3 p4 g 4 0 0 (7)

1 2 0 .5 p 2 s 2 = 0 3 4 0

2 p2 s2

2 0.5p2 s2 0 0 0
F

B. PCT in Control Analysis The addition of control laws to a serial manipulator simply changes the right-hand side of (6). For example, using independent potential difference (PD) control on the joints causes no issues when applying PCT. Nonlinear control laws can be more vexing if the nonlinearity is not polynomial or trigonometric based. Care must be exercised to ensure that additional nonlinearities are not introduced by division as well.

where c2 and s2 are cos(2 ) and sin(2 ), respectively, and


4

p1 =
i= 1

2 2 2 2 Ii + m 1 x2 1 + m 2 (x2 + a1 ) + (m 3 + m 4 )(a1 + a2 )

p2 = 2 [a1 x2 m2 + a1 a2 (m3 + m4 )] III. PCT APPLIED TO SCARA ROBOT In this paper, we will assume a 4-DOF SCARA-type manipulator having variation in the mass (and subsequently, the inertia) of both the rst two links as well as payload, as shown in Fig. 1 and Table I. Variation in the link lengths and centers of mass could also be incorporated into the model but are left out as their effects are assumed to be small compared to the mass variation. The other mechanical parameters are set to reasonable values, as given in Table I. The variation is assumed to be uniformly distributed across the range.2 p3 =
i= 2 4 2 Ii + m 2 x2 2 + a 2 (m 3 + m 4 )

p4 = m 3 + m 4 p 5 = I4 (8)

where i is the input torque (or force), Ii is the moments of inertia around the centroid, mi is the mass, xi is the mass center, and ai is the length for link i. In order to include the affect of mass on the inertia, the inertia terms are further dened as Ii = m i 2 i, i = 1, . . . , 4 (9)

A. Dynamics of the SCARA Robot An open kinematic chains dynamics can be derived utilizing either a NewtonEuler or Lagrangian formulation. Utilizing a Lagrangian formulation, the equations of motion for the SCARA robot as shown

where i is the radius of gyration of the links. Assuming a PD control law on each joint, the torques are i d e s i ), i = K p i (i d e s i ) + K d i ( i = 1, . . . , 4 (10)

should be noted that virtually any type of distribution and even mixed distributions (one variable being normally distributed and one being uniform) can be incorporated into the framework.

2 It

where des denotes the desired values. In order to use a general nonlinear ordinary differential equation solver, the equations are expressed in state space form [23]. For this particular problem, the governing equations can be put into eight

Authorized licensed use limited to: RWTH AACHEN. Downloaded on February 18, 2009 at 03:05 from IEEE Xplore. Restrictions apply.

208

IEEE TRANSACTIONS ON ROBOTICS, VOL. 25, NO. 1, FEBRUARY 2009

rst-order equations with

1 u1 u 2 2 u 3 3 u 4 4 u = u5 1 u 6 2 u 7 3 u8 where = u and M and F are dened in (7). B. SCARA Dynamic Equations Incorporating Uncertainty u5 u6 u7 u8 M 1 F 4

TABLE II PD CONTROLLER GAINS FOR NUMERICAL STUDY

(11)

(12)

Now that all the nonlinearities have been eliminated, the equation can be projected onto the basis. These inner products are cumbersome to compute as they contain a large quantity of terms. In order to avoid mistakes, these were computed using an automated script in MAPLE [25]. Due to the projection, four equations are created for every row in (12) for a total of 32 equations. These 32 equations are now integrated using a standard ODE solver (in this case, MATLABs ode45 function was utilized) to solve for the 32 unknown PC coefcients (i.e., ui j ) as functions of time. The time history of these can then be substituted back into (13) to determine the mean and variation. Due to the uncertain variables i in (13), a mini-MC is performed on the output. An analytical solution to the problem has also been proposed in [26]. IV. RESULTS In order to show the power of the PCT formulation, the SCARA robot was put under closed-loop control with the parameters shown in Table II. The rst two joints were commanded to go to 30 and 20 (i.e., u1 d e s = 0.5236 rad and u2 d e s = 0.3491 rad). Controller gains were chosen to show differing types of output (e.g., overdamped versus underdamped, large variation versus small variation). The results of this analysis are shown in Fig. 2. An MC analysis was also performed to compare with the solution. Both the mean and plus/minus three standard deviations (3 ), were plotted to show the resulting variation. A normal plot (not shown) was created at several time intervals to observe the normality of the output distribution. The normal plot showed that the distributions are normal with shortened tails. Thus, the 3 curves represent more of the distribution and slightly overestimate the variation that does not cause a problem for this study. Analysis of the plots show that the PCT analysis allows for the dynamic analysis of the stochastic differential equations of motion, and in this case, yields nearly identical results to the MC simulation. For this particular problem, the difference is shown in Table III. The most drastic difference is shown in joint 3 where some noise is seen in the PCT formulation. More accurate results can be garnered by adding more terms on the PCT expansion [27]. Also of interest is that the transition between underdamped and overdamped oscillations in 2 are handled easily. Also, the amount of variation if large or small is easily accommodated by the framework as shown in the graph for 3 . PCT has traded the number of ODE solver calculations by expanding the number of differential equations from 8 to 32 and requiring the calculation of 32 inner products, 16 of which were trivial. Once the inner products are determined, the analysis takes signicantly less time as compared to the MC analysis. For a sample size of 1000, the MC simulation took approximately 159.70 s to compute while the PCT method took a mere 1.27 s. Thus, scenario analyses can be performed easily in real time without having to run numerous datasets. It is interesting to note the existence of times where the total variation decreases signicantly. Further numerical analysis shows that at these points, the percentage error in the standard deviation estimate is the greatest. These situations (here called variation nodes) could possibly be useful for system testing and evaluation. The use of PCT in the controller design allows control analysis quicker than a traditional MC. Different controller gains can easily be tried and the resulting dynamics can be simulated. The PCT procedure

The governing equations can then be expanded using the PCT framework. The benet of using PCT is that it allows one to solve the stochastic differential equations with an expanded differential equation set. In other words, the random distributions need to be calculated only once and the ODE solver needs to run only once. This dramatically improves the efciency of the solver. The tradeoff is that the number of differential equations as well as the calculation of numerous inner products are much larger. Thus, as the number of unknown variables becomes large, the method becomes cumbersome, and a traditional MC would be better [24]. However, once the basis is chosen, these inner products need to be calculated only once. The rst step is to express all our variables in the appropriate basis. For this particular example, a Legendre polynomial basis is chosen due to the uniform distribution on the inputs.3 Each state space variable (ui ) is expanded in terms of the corresponding PC coefcients (ui j ) u i = u i 0 (t) + u i 1 (t)1 + u i 2 (t)2 + u i 3 (t)3 i = 1, . . . , 8. (13)

To keep the notation simpler, the use of the hat is omitted and the dependence of the polynomial coefcients on time will be dropped. For simplicity, the polynomial basis is truncated at rst order. As will be shown in this example, the results will concur with this assumption. We can now proceed to substitute the approximations into (12). The difculty is that the equations contain trigonometric functions. These trigonometric functions will cause problems with the methodology and must be expressed as a linear combination of the stochastic variables 1 , 2 , and 3 . Using a Taylors series approximation method as in [11], one can approximate the trigonometric functions as cos u2 cos u2 0 sin u2 0 (u2 1 1 + u2 2 2 + u2 3 3 ) sin u2 sin u2 0 + cos u2 0 (u2 1 1 + u2 2 2 + u2 3 3 ) (14)

where a rst-order polynomial approximation is utilized in this case. Higher order approximations will not cause problems with the formulation.
3 It should be noted that the use of the Legendre may not be the most appropriate as the output variations may be better represented by a Gaussian distribution, and thus, a Hermite basis may be better [10].

Authorized licensed use limited to: RWTH AACHEN. Downloaded on February 18, 2009 at 03:05 from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON ROBOTICS, VOL. 25, NO. 1, FEBRUARY 2009

209

Fig. 2.

Simulation results for all four joints of a SCARA robot with variation in the rst two link masses and the payload. TABLE III MAXIMUM ABSOLUTE DIFFERENCE BETWEEN PCT AND MC FOR SCARA ROBOT

6) PCT has no problem with both underdamped and overdamped type responses. Additionally, this example also corroborates several of the known advantages regarding the use of PCT. Specically, PCT is more efcient from a simulation time standpoint, allows dynamic changing of parameters during the simulation, and yields a richer result that can be utilized in other applications like controller design. REFERENCES
[1] D. Hoeltzel and W.-H. Chieng, A unied approach to the kinematic analysis of joint clearances and link length tolerances for determination of the rotational and positional accuracy of planar mechanisms, in Proc. Adv. Des. Autom. Conf. (ASME Des. Eng. Div.), Montreal, QC, Canada, Sep. 1989, vol. 19-3, pp. 345356. [2] J. Rhyu and B. Kwak, Optimal stochastic design of four-bar mechanisms for tolerance and clearance, J. Mechanisms, Transmiss., Autom. Des., vol. 110, no. 3, pp. 255262, Sep. 1988. [3] S. Lee and B. Gilmore, The determination of the probabilistic properties of velocities and accelerations in kinematic chains with uncertainty, J. Mech. Des., vol. 113, no. 1, pp. 8490, Mar. 1991. [4] P. Voglewede and I. Ebert-Uphoff, Application of workspace generation techniques to determine the unconstrained motion of parallel manipulators, ASME J. Mech. Des., vol. 126, no. 2, pp. 283290, Mar. 2003. [5] B. Fischer, Mechanical Tolerance Stackup and Analysis. New York: Marcel Dekker, 2004. [6] F. Chiesi and L. Governi, Tolerance analysis with eM-TolMate, J. Comput. Inf. Sci. Eng., vol. 3, pp. 100105, Mar. 2003. [7] Z. Shen, Tolerance analysis with EDS/VisVSA, J. Comput. Inf. Sci. Eng., vol. 3, no. 1, pp. 9599, 2003. [8] Z. Shen, G. Ameta, J. Shah, and J. Davidson, A comparative study of tolerance analysis methods, J. Comput. Inf. Sci. Eng., vol. 5, no. 3, pp. 247256, Sep. 2005. [9] N. Wiener, The homogeneous chaos, Amer. J. Math., vol. 60, pp. 897 936, 1938. [10] D. Xiu and G. Karniadakis, The WienerAskey polynomial chaos for stochastic differential equations, SIAM J. Sci. Comput., vol. 24, no. 2, pp. 619644, 2003. [11] B. J. Debusschere, H. N. Najm, P. P. Pebayt, O. M. Knio, R. G. Ghanem, and O. P. Le Maitre, Numerical challenges in the use of polynomial chaos

also opens the door for easy optimization of the controller gains for the small amount of variation. Sensitivity studies can quickly and effectively be performed for different controller gains. V. CONCLUSION This paper has shown in detail how to use PCT to analyze the dynamic response of an open-loop mechanism by applying it to a SCARA robot manipulator. Through this particular example, several items were found. 1) PCT on the SCARA robot is feasible as long as the number of unknowns are small. Automating the process further aiding in speeding up the process. However, this even has its limits as even the automated process is still cumbersome. 2) Using PCT on robotic applications requires a judicious choice of states and formulation of the problem to make sure that nonlinearities are not introduced into the equations. 3) PCT on a SCARA robot gives consistent results to a large MC for a simple one-term expansion. This was true even using a standard approximation of the trigonometric identities. 4) PCT exhibits antinodes (positions in the output where the variation decreases signicantly), even in this robotics example. However, this was not noted on all joints. 5) PCT can be utilized with feedback control. However, integral and nonlinear control poses some interesting problems that have yet to be solved.

Authorized licensed use limited to: RWTH AACHEN. Downloaded on February 18, 2009 at 03:05 from IEEE Xplore. Restrictions apply.

210

IEEE TRANSACTIONS ON ROBOTICS, VOL. 25, NO. 1, FEBRUARY 2009

[12] [13] [14] [15] [16] [17] [18] [19]

representations for stochastic processes, SIAM J. Sci. Comput., vol. 26, no. 2, pp. 698719, 2005. X. Wan and G. E. Karniadakis, Beyond WienerAskey expansions: Handling arbitrary PDFs, J. Sci. Comput., vol. 27, no. 13, pp. 455464, Jun. 2006. D. Xiu, D. Lucor, G. E. Su, and C.-H. Karniadakis, Stochastic modeling of ow-stucture interactions using generalized polynomial chaos, J. Fluids Eng., vol. 124, pp. 5159, Mar. 2002. D. Lucor, C.-H. Su, and G. Karniadakis, Generalized polynomial chaos and random oscillators, Int. J. Numer. Methods Eng., vol. 60, no. 3, pp. 571596, May 2004. D. Xiu and G. Karniadakis, Modeling uncertainty in steady state diffusion problems via generalized polynomial chaos, Comput. Methods Appl. Mech. Eng., vol. 191, pp. 49274948, 2002. A. Monti, F. Ponci, and T. Lovett, A polynomial chaos theory approach to uncertainty in electrical engineering, in Proc. Int. Conf. Intell. Syst. Appl. Power Syst., Arlington, VA, Nov. 2005, pp. 534539. A. Sandu, C. Sandu, and M. Ahmadian, Modeling multibody dynamic systems with uncertainties. Part I: Theoretical and computational aspects, Multibody Syst. Dyn., vol. 15, no. 4, pp. 369391, May 2006. A. Sandu, C. Sandu, and M. Ahmadian, Modeling multibody dynamic systems with uncertainties. Part II: Numerical applications, Multibody Syst. Dyn., vol. 15, no. 3, pp. 241262, Apr. 2006. L. Li, C. Sandu, and A. Sandu, Modeling and simulation of a full vehicle with parametric and external uncertainties, in Proc. ASME

[20] [21] [22]

[23] [24] [25] [26]

[27]

Int. Mech. Eng. Congr. Expo., Nov. 2005, vol. 118A, no. 1, pp. 209 214. J. Craig, Introduction to Robotics: Mechanics and Control, 3rd ed. Reading, MA: Addison-Wesley, 2005. L.-A. Dessaint, M. Saad, and B. H ebert, K. Al Haddad, An adaptive controller for a direct-drive SCARA robot, IEEE Trans. Ind. Electron., vol. 39, no. 2, pp. 105111, Apr. 1992. L.-A. Dessaint, M. Saad, and B. H ebert, C. Gargour, An adaptive controller for a direct-drive SCARA robot: Analysis and simulation, in Proc. 16th Annu. Conf. IEEE Ind. Electron. Soc., Pacic Grove, CA, Nov. 1990, vol. 1, pp. 414420. J. Ginsberg, Mechanical and Structural Vibrations: Theory and Applications. New York: Wiley, 2001. P. Prempraneerach, Uncertainty analysis in a shipboard integrated power system using multi-element polynomial chaos, Ph.D. dissertation, Massachusetts Inst. Technol., Cambridge, MA, 2007. A. Smith, A. Monti, and F. Ponci, Robust stability and performance analysis using polynomial chaos theory, in Proc. Soc. Comput. Simul. Summer Conf., Jul. 2007, pp. 4552. G. DAntona, A. Monti, F. Ponci, and L. Rocca, Maximum entropy analytical solution for stochastic differential equations based on the WienerAskey polynomial chaos, IEEE Trans. Instrum. Meas., vol. 56, no. 3, pp. 689695, Jun. 2007. R. Field and M. Grigoriu, On the accuracy of the polynomial chaos approximation, Probablistic Eng. Mech., vol. 19, pp. 6580, 2004.

Authorized licensed use limited to: RWTH AACHEN. Downloaded on February 18, 2009 at 03:05 from IEEE Xplore. Restrictions apply.

You might also like