Вы находитесь на странице: 1из 7

2013 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM) Wollongong, Australia, July 9-12, 2013

Reducing the Optimization Problem for the Efcient Motion Planning of Kinematically Redundant Parallel Robots
Sebastian Niemann1 , Jens Kotlarski2 , Tobias Ortmaier2 , and Christian Mller-Schloer1

Abstract This paper presents an optimization procedure used to efciently maximize the potentials of parallel manipulators with kinematic redundancy within real-time. The proposed approach consists of reducing the search space dened by the optimization problem beforehand, through minimizing the amount of optimization points necessary to induce the optimization problem. Furthermore, the computation time of the tness function is reduced during run-time. Therefore, the relationship between the Cartesian path of the end-effector and the resulting optimization problem is studied and a procedure to minimize the given dependencies is presented. Exemplarily, a kinematically redundant 3(P)RRR parallel robot is considered to quantify the efciency of the discussed procedure. The results demonstrate that the proposed approach is able to outperform existing procedures by one to two orders of magnitude.

I. I NTRODUCTION In comparison to classical serial mechanisms parallel kinematic machines (PKM) provide higher accuracy, higher stiffness, and better dynamic properties [1]. However, a main drawback of such mechanisms is the small workspace to installation space ratio, which is further reduced by typetwo singularities occurring right within the workspace [2]. In order to minimize the singularity loci of parallel mechanisms and to increase their performance redundancy can be used [3]. Two redundancy approaches are established for PKM, actuation redundancy and kinematic redundancy [4], [5]. Here, kinematic redundancy is proposed, realized by adding at least one actuated joint to one kinematic chain [6]. Thanks to the additional actuator(s), the inverse displacement problem has an innite number of solutions [7]. Hence, recongurations of the mechanism can be performed selectively in order to avoid singularities and to affect its performance directly [8], [9], [10], [11]. In order to achieve and to maximize the aforementioned potentials of kinematic redundancy, an appropriate optimization of the position of the additional actuator(s) is required. In [10] different optimization strategies are presented. They are based on two main approaches: A selective discrete optimization and a classical continuous optimization. Analysis examples demonstrate the efciency, in terms of increasing
1 S. Niemann and C. Mller-Schloer are with the Institute of Systems Engineering, System and Computer Architecture, Leibniz Universitt Hannover, Hanover, Germany. {niemann,cms}

@sra.uni-hannover.de
2 J. Kotlarski and T. Ortmaier are with the Institute of Mechatronic Systems, Leibniz Universitt Hannover, Hanover, Germany. {jens.kotlarski,tobias.ortmaier}

performance, of kinematic redundancy with respect to the introduced optimization procedures. Beyond, in [11] several experiments are performed, validating the potential of kinematic redundancy in practice. In addition, optimization strategies for the motion planning of kinematically redundant PKM as well as optimization criteria have already been discussed in several publications, like [8], [10], [11], [12], [13], [14]. The same holds true for the concept of kinematic redundancy itself, see e. g. [6], [7], [10], [11]. Hence, the capability of kinematic redundancy, i. e. the potential of additional prismatic actuators, is not addressed in this paper. Several optimization algorithms to solve the resulting optimization problem, have been applied [15], [16], [17], with particle swarm optimization (PSO) being the most suited algorithm [18], [19]. However, none of these approaches, combined or used alone, is able to provide a satisfactory solution for the optimization problem, i. e. online path planning, within real-time1 . To reduce the time-consuming aspect of the optimization task, Barbosa et al. [20] and Gao et al. [21] used an articial neural network to approximate the calculation of the optimization criteria, achieving a remarkable reduction of the search time by 30% to 50%. However, they would not achieve the optimization task within real-time. Therefore, this paper presents a procedure to solve the optimization task within real-time by minimization of the search space. It is based on determining the minimal number of optimization points needed to dene the objective function induced by any path of the end-effector. Note that the path of the end-effector is not modied. A second approach combines the iterative behavior of optimization algorithms with the calculation of the tness function, obtaining an additional termination condition. Hereby, we outperform existing optimization procedures by one to two orders of magnitude. Note that we only reduce the optimization problem, but do not propose a new optimization algorithm or optimization strategy for the motion planning of kinematically redundant PKM. Sec. II briey reviews the fundamentals of optimization problems and introduces the optimization criterion addressed in this paper. In Sec. III the proposed optimization procedure is elucidated. Therefore, the causal relation between the motion planning and the resulting optimization problem is reviewed to minimize the number of dependencies, i. e.
1 A computing time below 100 ms is considered within real-time for the given application class.

@imes.uni-hannover.de

978-1-4673-5320-5/13/$31.00 2013 IEEE

618

the number of optimization points. Furthermore, the iterative behavior of optimization algorithms is embedded into the tness function. Sec. IV demonstrates the capability of the proposed optimization procedure. Exemplarily, a kinematically redundant 3(P)RRR PKM is considered in order to quantify the efciency of the motion planning. After a brief description of the mechanism the results of the optimization procedure are presented and discussed. Sec. V concludes this paper and addresses future work. II. O PTIMIZATION P ROBLEM An optimization problem is the problem of nding the best acceptable solution within the search space. Formally, an optimization problem is dened by the search space S , an objective function F and a goal. Commonly, the goal is to nd the minimal or maximal value of the objective function, with respect to the optimal parameter x S , within a reasonable amount of time. By convention, the optimization problem shall be viewed as a maximization problem: maximize F (x), F : S R.
xS

It is important to note that many alternative performance criteria can be used to optimize the mechanisms conguration. Without loss of generality, the exemplarily chosen 2-norm condition number, i. e. its reciprocal called dexterity [14], of the homogenized Jacobian is considered. III. O PTIMIZATION P ROCEDURE As stated before, optimization strategies as well as optimization criteria have already been discussed in several publications. In contrast, to the best of the authors knowledge, an optimization procedure to efciently nd an appropriate solution for the motion planning of kinematically redundant PKM has not been addressed yet. In most cases only the suitability of optimization algorithms is considered, without being able to present a satisfactory solution to the optimization problems for real-time operations, i. e. online path planning. With respect to (1) the reduction of computation time needed to solve the optimization problem can be achieved by minimizing the complexity of the search space or of the objective function. Both can be done beforehand or during the optimization task. An optimization algorithm, e. g. PSO, reduces the search space at run time by analyzing a subset of the search space in regard to the current solution [25], [26]. The approximation of the objective function by applying an articial neural network, described by Barbosa et al. [20] and Gao et al. [21], is a helpful approach to lower the complexity of the objective function in advance. However, this alone is not sufcient to achieve the considered optimization task within real-time. Therefore, this paper presents a procedure to minimize the search space beforehand, as well as speed up the calculation of the objective function during run time. The minimization of the search space is based on providing a minimal number of optimization points needed to dene the objective function induced by any path of the end-effector. Our second approach is to embed the iterative behavior of optimization algorithms into the objective function, giving an additional termination condition. Both approaches are then combined. Note that this is achieved independently from the optimization strategy. A. Minimizing the necessary optimization points In reference to (3), minimizing the optimization points results in minimizing the number of optimization criterion evaluations. Therefore, the relationship between the optimization points O and the path X E of the end-effector is reviewed. According to the current state of the art, the optimization points used in the denition of the objective function are identical to the trajectory points used to move the end-effector along a given path [10]. The discretization is thereby motivated by the impossibility to pointwise evaluate any continuous model. To overcome the information gap between neighboring optimization points caused by the discretization, an area A(o) Rn surrounding each point o with nearly the same property regarding the optimization criterion is introduced. The elements of the area are denoted as o .

(1)

In order to achieve and to maximize the aforementioned potentials of kinematic redundancy, an appropriate optimization of the positions q r Rm of the additional redundant actuators is required [22]. Therefore, the search space S of the optimization problem depends on a set O of optimization points o Rn , e. g. end-effector poses, and the positions q r of the additional actuators: S (Rn )l Rm , l = |O |. (2)

The induced objective function F is dened by the minimal value of each pair of an optimization point and the position q r , calculated by an optimization criterion f : F : (O , q r ) min{f (o1 , q r ), . . . , f (ol , q r )}, f : Rn Rm R. (3) (4)

Note that if the position q r is continuously adjusted, only a single optimization point oi O is considered in each time-step, thus l equals 1 in this case. The optimization can be realized with respect to several criteria [14]. In this paper, since several performance indices are directly related to the singularity loci [22], the authors focus exclusively on the avoidance of singularities. The 21 norm condition number (J h ) = J h 2 J 2 of the h 1 homogenized overall Jacobian matrix J h = A h B h is chosen to maximize the distance to the most relevant typetwo singularities with respect to a given Cartesian pose xE and a desired path X E , respectively2 , whereas its discrete representation is denoted X D E . A mechanism is in a singular conguration of type-two if (J h ) tends to innity. Hence, the optimization criterion to be maximized results to: q r = arg max ((J h ))
qr 1

(5)

2 The Jacobian J is not homogeneous in terms of physical units. Hence, a homogenization (J J h ), e. g. based on a unit length, is required in order to obtain appropriate values for (J h ) (see [23], [24] for details).

619

It is bounded by a feasible small deviation . o A(o) : |f (o , q r ) f (o, q r )| . (6)

With the shape of the area being usually complex, in the following the largest area C (o, o ) A(o) enclosed by a circle with radius o and center o is studied. Note that the motivation to study circles is only based on their simple description but did not limit the proposed procedure. In summary, the property, i. e. the optimization criterion, of a given path can be completely described, if the area C (o, o ) of all optimization points overlays the path: XE C (o, o ),
oO

o O : C (o, o ) X E = . (7)

Exemplarily, the basic idea is demonstrated in Fig. 1, showing a rectangular and an octagonal path. For simplication, assume that only the edge points are elements of XD E . By using the trajectory points as optimization points in Fig. 1a, their number is already minimal, since there is no overlaying using a reduced number, such that (7) holds. In contrast, the same does not apply for Fig. 1b, since Fig. 1c presents an overlay with only three optimization points. With respect to (7) it is not necessary that o is an element of the path X E , Fig. 1d gives an alternative solution.

optimization points given in Fig. 1a can also be used for the octagonal path. Finding the smallest amount of optimization points overlaying a given path in regard to (7) is a time consuming task, not being feasible within real-time. However, it is possible to estimate the largest area C (o, o ) around each optimization point beforehand, i. e. without considering any specic path X E . For this purpose, the workspace of the robot is represented by a grid O W , whereby the distance between neighboring points equals the minimal distance between trajectory points within X D E . For each point o within the grid, the maximal deviation between o and any element within a grid-based representation C (o, o ) is calculated, increasing the radius o as long as the maximal deviation is within the tolerance , as dened in (6). The set Qr of actuator positions used for calculating the deviation shall be randomly generated. Alg. 1 gives a pseudo implementation of this task. Algorithm 1: Determining the largest area C (o, o ) around each optimization point o O W = CreateGrid() Qr = CreateSamples 0 = 0 1 = foreach o O W do max = 0 i=1 repeat // Calculate maximal deviation foreach o C (o, i ) \ {o} do foreach q r Qr do max = max(max , |f (o , q r ) f (o, q r )|) end end i+1 = inc(i ) i=i+1 until max > o = i2 end It is important to note that the radius o equals zero if the maximal allowed deviation is exceeded for o = . In this case, the discretization of the trajectory X D E could not fully describe the optimization problem in the rst place. Based on this, the set O W of optimization points to overlay the workspace is minimized. Therefore, all remaining points o O W \ O W are ordered by their number of covering points o in respect to the already covered points O covered from previous iterations. The optimization point omax , covering the largest currently uncovered area, is considered necessary. Each necessary point is added to the set O W and all points within C (omax , omax ) are marked as covered. To speed up the process, the area o is calculated for each point o beforehand and only updated if the distance between the necessary and another point |o omax | is not greater than the sum of their radii o + omax . In addition, points

(a) Rectangular path with minimal number of optimization points

(b) Octagonal path with nonminimal number of optimization points

(c) Octagonal path with minimal number of optimization points

(d) Octagonal path with alternative placement of optimization points

Fig. 1: Demonstration of the basic idea, given a rectangular and octagonal path; the area enclosed by the dotted circle around each optimization point o equals C (o, o ) It is important to note, that a more complex path in the same area does not need a higher number of optimization points in general. More generally, the number of optimization points needed depends only on the area described by the path to be overlaid and its location within the workspace. This will be demonstrated later on in Sec. IV. Furthermore, different paths can induce the same optimization problem. E. g. the

620

that did not cover any uncovered area, i. e. o = 0, are not considered anymore. Alg. 2 gives a pseudo implementation of this task. Note that the proposed optimization procedure did not depend on the optimization criteria. Algorithm 2: Determining all necessary optimization points to overlap the workspace max = 0 O covered = O W = foreach o O W do o = |C (o, o )| max = max(max , o ) end while |O overlapped | < |O W | do // Find o with maximal o omax = 0 foreach o in O W \ O W do if o > omax then omax = o end end O W = O W {omax } foreach o C (omax , omax ) do O covered = O covered {o} end // Update o if needed foreach o C (omax , omax + max ) do if |o omax | o + omax then o = |C (o, o ) \ O covered | if o = 0 then O W = O W \ {o} end end end end

Fig. 2: Chosen optimization points for an exemplary path; The optimization points within a solid circle were chosen by the procedure described in Sec. III-B

optimization algorithm is called iterative, if a new solution is only accepted if the objective value is improved. This is a basic characteristic of most optimization algorithms [26]. Therefore, an iteration can be terminated prematurely if f (o, q r , i) < f (o, q r , i 1) holds. Based on (6), the optimization point with the worst objective value in the last iteration is likely to have the worst objective value again. Therefore, the objective values f (o, q r , i) are computed in the order of the previous objective values, starting with the worst. This signicantly reduces the computation time for cases where the optimization algorithm tends to evaluation inferiority solutions. The improvement however depends on the amount of optimization points used at each iteration. IV. E XAMPLE In the following, the capabilty of the proposed optimization procedure is demonstrated. Therefore, the calculation time as well as the number of function evaluations are compared with respect to the proposed improvements and for different numbers r of additional prismatic actuators, i. e. different dimensions of the optimization problem. Exemplarily, the motion planning of a kinematically redundant 3(P)RRR mechanism, being briey reviewed in Sec. IV-A, is considered. It is important to note that the capability of kinematic redundancy, i. e. the potential of additional prismatic actuators, as well as the PSO algorithm, are not addressed in this paper. A. Kinematically redundant 3(P)RRR mechanism Besides others, the planar, kinematically redundant 3(P)RRR3 mechanism (see Fig. 3) is introduced in [7] and [27]. Basically, it is similar to the non-redundant 3RRR mechanism studied amongst others in [28]. Three kinematic chains Gi Mi Pi (i = 1, 2, 3) connect the moving platform P1 P2 P3 to the base G1 G2 G3 . In order to achieve kinematic redundancy, at least one prismatic actuator is added to the structure. This allows to linearly move the position r Gi = (xGi , yGi )T of at least one active base joint Gi with respect to the actuator position i . Depending on the degree of redundancy r , the mechanism is driven by 3 + r actuators. In case of a maximum number
3 The bracketed P indicates that not necessarily each kinematic chain contains an additional prismatic actuator.

B. Selecting the optimization points The selection of the optimization points during run time is based on Alg. 2. Instead of overlapping the workspace, only the path of the end-effector needs to be overlapped. Therefore, only optimization points that cover at least a part of the path are considered, i. e. O W is replaced by {o D O W : C (o, o ) X E = }. Since we want the path to be covered, the condition of the while-loop needs to be replaced by |O overlapped | < |X D E | and C (o, o ) needs to be replaced by C (o, o ) X D E , except for C (omax , omax + max ) since this loops over all optimization points to be updated. O W will then denote the minimal number of optimization points for the optimization problem (3). The result of this approach are shown in Fig. 2 for an exemplary path. C. Embedding the iterative behavior In addition to the minimization of the search space, the exploitation of the iterative behavior of optimization algorithms allows for a dynamic reduction of the objective function. An

621

tolerance = 0.05. It can be clearly seen, that the maximal radius increases near the center of the x-axis (x = 0.6 m), reaching the overall maximal radius r = 0.0355 m at the lower part (x = 0.6 m, y = 0 m). In contrast, as mentioned in Sec. III-A this case shows areas within the workspace with radius r = 0 m assigned, meaning that the smallest considered area C (o, o ) does not fulll (6). Looking at the distribution of singularity loci given in [10], the mentioned areas are in the same region where singularity loci of the non-redundant mechanism need to be avoided. Further on, only the areas with a radius assigned are examined. Fig. 3: Kinematically redundant 3(P)RRR mechanism with a maxmimum number of three additional prismatic actuators
1 0.8 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 Workspace x-axis [m] 1 1.2 0 1 102 [m] 3 102

Workspace y-axis [m]

of three additional prismatic actuators (r = 3), the system input is given by qa = ,


T

qT r

= (1 , 2 , 3 , 1 , 2 , 3 ) .

(8)

2 102

The conguration of the moving platform, i. e. the system output, is dened by xE = (xE , yE , E ) ,
T

(9)

where xE and yE represent the point of origin of (CF)E and (0) E is its orientation around ez . Due to space limitations, the inverse kinematic model = f (xE , q r ) as well as the denition of the Jacobians A and B , i. e. the Jacobians of the direct and of the inverse kinematics, are not presented here. A complete description can be found in [29]. According to the 3(P)RRR prototype presented in [11], the geometrical parameters of the exemplarily considered mechanism are given in TABLE I. Here, the orientation of the TABLE I: Geometrical parameters of the exemplarily considered 3(P)RRR mechanism; entries may vanish with respect to r
(0) xGi [m] (0) yGi

Fig. 4: Maximal radius o of additional prismatic actuators, with starting distance = 0.005 m and tolerance = 0.05 Based on the results mentioned above, Fig. 5 displays the placement of the minimal number of optimization points used to overlay the workspace, computed using Alg. 2. Since the placement is directly related to the maximal radius, the gaps between the optimization points are larger close to the middle of the workspace, being more dense close to the boundary of the workspace.

[m]

(E) xPi [m] (E) yPi [m]

li,1 li,2 i i i

[m] [m] [ ] [m] [m]

i=1 0 .6 1.08 0 .0 0.12/3 0 .6 0 .6 0 .0 0 .5 0.5

i=2 0.0 0.0 0 .1 0.03/3 0.6 0.6 90.0 0.2 0.8

i=3 1.2 0.0 0 .1 0.03/3 0.6 0.6 90.0 0.2 0.8


(0)

prismatic actuator i with respect to the ex -axis is denoted by i and i and i are its maximal and minimal stroke, respectively. B. Results Regarding the 3(P)RRR mechanism with its maximal number of additional prismatic actuators r = 3, Fig. 4 displays the maximal radius o for each optimization point, obtained using Alg. 1 with a starting distance = 0.005 m and a Fig. 5: Distribution of the minimal number of optimization points used to overlay the workspace, based on Fig. 4 The evaluation uses PSO, especially SPSO-11 [30], as

622

Evaluations of optimization criterion (#FEs)

optimization algorithm and measures the number of function evaluations (#FEs) as well as the computation time needed to obtain the optimal solution regarding the optimization criteria, i. e. condition number (J h ) (see (5)). As an overall termination condition, the optimization algorithm stops after 50.000 evaluations of the optimization criterion. There are 50 optimization problems given, induced by a circular path with radius 0.175 m, placed randomly within the workspace of the 3(P)RRR PKM. Note that a circular path is chosen since it is not biased in favor of the presented procedure4 . Since PSO uses random values within the optimization algorithm, each optimization problem is solved 30 times, giving the mean value as result. The computing time is based on Matlab R , using a Lenovo R T420s Notebook with an Intel R i7-2640M, 2.80 GHz CPU. Note that using a more sophisticated base like C\C++ instead of Matlab R the resulting computation time may be considerably lower. Since the classical continuous optimization denes an easier problem, as mentioned in Sec. II, a selective discrete optimization strategy is examined. To clarify that the optimization function FO induced by the minimal number of essential optimization points O equals the optimization function FX D induced by O = X D E E the best solutions q r,O , q r,X D of the corresponding optiE mization function are compared. Fig. 6 shows the difference between FX D (X D ) and FX D (X D E , q r,X D E , q r,O ), where a E E E positive value is in favor of the stated procedure. Besides a few outliers, the best solution achieved by the proposed procedure is always within the overall best solutions. Surprisingly, the obtained solutions are even better than optimizing FX D directly. This can be explained by the imposed maxiE mum number of 50.000 evaluations per solution. Therefore, the reduced number of optimization points results in a higher convergence rate.
Deviation of tness values 0.3 0.2 0.1 0 0 10 20 30 40 50 Optimization problem, i. e. circular path

points. Notably, using the trajectory points, the number of evaluations does not drop below 1.000 iterations, while using the minimized optimization points, 1.000 iterations are only slightly exceeded at most (see Fig. 7). TABLE II: Average results of the evaluation over all 50 optimization problems for each possible number r of additional prismatic actuators
#FEs 10175 8327 374 304 r = 1 time [ms] 1545.98 1265.20 56.83 46.19 #FEs 8844 7279 415 341 r = 2 time [ms] 1343.75 1105.97 63.10 51.81 #FEs 10108 7701 732 593 r = 3 time [ms] 1535.80 1170.10 111.22 90.10

u/s c/s u/m c/m

Embedding the iterative behavior results in a small reduction of the computation time only. This can be explained by the used optimization algorithm, which has a high rate of guessing a better solution at each iteration. In contrast, the minimization of the optimization points reduces the number of evaluations up to two orders of magnitude, with respect to (u/s). Completing the evaluation, TABLE II lists the average number of needed evaluations as well as the computation time for each combination of approaches from one to three additional prismatic actuators of the 3(P)RRR PKM. While (u/s) is not able to solve the problems within 1.5 s on average, (c/m) on contrary solves the problem within 100 ms.

104

103

Fig. 6: Difference of the tness values of the optimal solution between optimization problems using the minimal number of optimization points and the trajectory points Additionally, the number of optimization criterion calculations and the required computing time is evaluated. Therefore, each optimization problem is solved beforehand, using the best obtained tness value as an additional termination condition. Every combination using the approaches described in Sec. III-A and Sec. III-C is evaluated. Thereby, FX D is denoted as (u/s), meaning usage of the unchanged E tness function with the trajectory points as optimization
4 The proposed procedure needs only one optimization point for each intersection, while a procedure based on the trajectory needs an optimization point for each time the path crosses itself.

20 30 40 50 Optimization problem unchanged tness function / standard points (u/s) changed tness function / standard points (c/s) unchanged tness function / minimized points (u/m) changed tness function / minimized points (c/m)

10

Fig. 7: Evaluation of number of function evaluations needed until reaching the predened optimal solution; in contrast to the dotted lines, the solid lines make use of the iterative behavior, whereas the lines with a circle mark the exploitation of the minimal number of optimization points

623

V. C ONCLUSION In this paper, a procedure to reduce the computation time of the motion planning of kinematically redundant PKM was presented. After a short description of the main idea including a denition of the given optimization problem, the proposed procedure was explained in detail. Basically, it includes a systematic minimization of the number of necessary optimization points as well as an exploitation of the iterative behavior of optimization algorithms. Exemplarily, a kinematically redundant 3(P)RRR with up to three additional actuators was considered to quantify the efciency of the procedure. The results demonstrate that the proposed approach is able to signicantly reduce the calculation time for nding appropriate positions of the additional (redundant) actuators. In comparison to classical optimization procedures, the calculation time decreases by one to two orders of magnitude. Furthermore, it was exemplarily demonstrated that a feasible solution can be obtained within real-time. However, optimization problems regarding more complex mechanisms with further degrees of freedom may not be solved within the assumed real-time limit. Therefore, future work will deal with reducing the computation time for higher search space dimensions by exploiting the small computing time for inferior solutions in combination with tabu-based algorithms. Here, the premature termination described in Sec. III-C is used to exclude areas with inferior tness values within the search space. Aside from this, based on analyzing only single optimization points beforehand, the prediction of an optimal parameter for a given path will be analyzed by calculating the crucial optimization points in reference to (3). R EFERENCES
[1] J.-P. Merlet, Parallel Robots (Second Edition). Springer, 2006. [2] C. M. Gosselin and J. Angeles, Singularity analysis of closed-loop kinematic chains, IEEE Trans. Robot. Autom., vol. 6, pp. 281290, 1990. [3] J.-P. Merlet, Redundant parallel manipulators, Laboratory Robot. and Automation, vol. 8, pp. 1724, 1996. [4] S. Kock and W. Schumacher, A parallel x-y manipulator with actuation redundancy for high-speed and active-stiffness applications, in Proc. 1998 IEEE Int. Conf. Robot. Autom., 1998. [5] J. Wang and C. M. Gosselin, Kinematic analysis and design of kinematically redundant parallel mechanisms, J. of Mech. Design, vol. 126, pp. 109118, 2004. [6] M. G. Mohamed and C. M. Gosselin, Design and analysis of kinematically redundant parallel manipulators with congurable platforms, IEEE Trans. Robot. Autom., vol. 21, pp. 277287, 2005. [7] I. Ebrahimi, J. A. Carretero, and R. Boudreau, 3-PRRR redundant planar parallel manipulator: Inverse displacement, workspace and singularity analyses, Mech. Mach. Theory, vol. 42, pp. 10071016, 2007. [8] I. Ebrahimi, J. A. Carretero, and R. Boudreau, Actuation scheme for a 6-dof kinematically redundant planar parallel manipulator, in Proc. 12th IFToMM World Congress, 2007. [9] J. Kotlarski, T. Do Thanh, H. Abdellatif, and B. Heimann, Singularity avoidance of a kinematically redundant parallel robot by a constrained optimization of the actuation forces, in Proc. 17th CISM-IFToMM Symp. on Robot Design, Dynamics and Control, 2008. [10] J. Kotlarski, T. Do Thanh, B. Heimann, and T. Ortmaier, Optimization strategies for additional actuators of kinematically redundant parallel kinematic machines, in Proc. 2010 IEEE Inter. Conf. on Robot. and Autom., 2010.

[11] J. Kotlarski, B. Heimann, and T. Ortmaier, Experimental validation of the inuence of kinematic redundancy on the pose accuracy of parallel kinematic machines, in Proc. 2011 IEEE Inter. Conf. on Robot. and Autom., 2011. [12] J. Carretero, I. Ebrahimi, and R. Boudreau, A comparison between two motion planning strategies for kinematically redundant parallel manipulators, in Advances in Robot Kinematics: Analysis and Design. Springer, 2008. [13] R. Boudreau and S. Nokleby, Force optimization of kinematicallyredundant planar parallel manipulators following a desired trajectory, Mech. Mach. Theory, vol. 56, pp. 138155, 2012. [14] J.-P. Merlet, Jacobian, manipulability, condition number, and accuracy of parallel robots, J. Mech. Design, vol. 128, pp. 199206, 2006. [15] I. E. Moghaddam, Kinematic redundancy in planar parallel manipulators, Ph.D. dissertation, University of New Brunswick, Canada, 2008. [16] T.-S. Zhan and C.-C. Kao, Modied pso method for robust control of 3rps parallel manipulators, Math. Problems in Eng., vol. 1, p. 25, 2010. [17] I. Tyapin, G. Hovland, and T. Brogardh, Workspace optimisation of a recongurable parallel kinematic manipulator, in IEEE/ASME Inter. Conf. Advanced Intell. Mech.,, 2007. [18] A. Mishra and S. N. Omkar, Singularity analysis and comparitive analysis of six degree of freedom stewart platform as a robotic arm by heuristic algorithms and simulated annealing. Inter. J. Eng. Sci. and Tech., vol. 3, pp. 644659, 2011. [19] Q. Xu and Y. Li, Stiffness optimization of a 3-dof parallel kinematic machine using particle swarm optimization, in IEEE Inter. Conf. Robot. and Biomimetics, 2006. [20] M. Barbosa, E. Pires, and A. Lopes, Optimization of parallel manipulators using evolutionary algorithms, in 5th Inter. Workshop Soft Comp. Models in Ind. and Env. Applicat. Springer, 2010. [21] Z. Gao, D. Zhang, and Y. Ge, Design optimization of a spatial six degree-of-freedom parallel manipulator based on articial intelligence approaches, Robot. Comput.-Integr. Manuf., vol. 26, pp. 180189, 2010. [22] J. Kotlarski, H. Abdellatif, and B. Heimann, Improving the pose accuracy of a planar 3RRR parallel manipulator using kinematic redundancy and optimized switching patterns, in Proc. 2008 IEEE Inter. Conf. Robot. and Automat., 2008. [23] L. Stocco, S. Salcudean, and F. Sassani, Matrix normalization for optimal robot design, in Proc. 1998 IEEE Inter. Conf. Robot. and Autom., 1998. [24] J. Angeles, Is there a characteristic length of a rigid-body displacement, Mech. Mach. Theory, vol. 41, pp. 884896, 2006. [25] J. Kennedy and R. Eberhart, Particle swarm optimization, in Proc. IEEE Inter. Conf. Neural Networks, 1995. [26] T. Weise, Global Optimization Algorithms: Theory and Application, 2nd Edition, T. Weise, Ed. Thomas Weise, 2009. [27] J. Kotlarski, H. Abdellatif, and B. Heimann, On singularity avoidance and workspace enlargement of planar parallel manipulators using kinematic redundancy, in Proc. 13th IASTED Inter. Conf. Robot. and Applicat., 2007. [28] C. M. Gosselin and J. Angeles, The optimum kinematic design of a planar three-degree-of-freedom parallel manipulator, J. of Mech. Transm. and Automat. in Design, vol. 110, pp. 3541, 1988. [29] J. Kotlarski, B. Heimann, and T. Ortmaier, Inuence of kinematic redundancy on the singularity-free workspace of parallel kinematic machines, Frontiers of Mech. Eng., vol. 7, pp. 120134, 2012. [30] M. Clerc, Standard particle swarm optimisation, Online, September 2012.

624

Вам также может понравиться