approaches。 The three evolutionary optimization approaches employed are Simple Genetic Algorithms, Genetic Algorithms with elitism, and Differential Evolution (DE)。 These approaches were used for the optimum design of SCARA and articulated type manipulators based on kinematic, dynamic and structural analyses。 The objective function minimizes the torque required for the motion subject to deflection and physical constraints with the design variables being the link physical characteristics (length and cross-sectional area parameters)。 In this work, we experimented with various cross-sections for the links。 The main findings of this research are that the DE converges quickly, requires significantly less number of iterations and achieves better results by reaching smaller  objective functions。72194

Keywords。  Robot design; Optimization; Evolutionary techniques; Dynamic; Kinematic;   Structural

1。 INTRODUCTION

In this article, we discuss  optimum  robot  design  based  on  task  specifications  using three evolutionary optimization approaches。 The three evolutionary optimization approaches employed are simple genetic algorithm (SGA), genetic algorithm with elitism (GAE), and differential evolution (DE)。 These approaches were used for the optimum design of SCARA and articulated type manipulators。 The objective function minimizes the torque subject to deflection and physical constraints for a  defined manipulator motion with the design variables being the link physical characteristics (length   and   cross-sectional   area   parameters)。   The   links    of    serial   manipulators are  usually   over  designed   in  order  to  be  able  to  support  the  subsequent   links      on

P。S。 SHIAKOLAS et at。

the chain and the payload to be manipulated。  However,  increasing  the  size  of the links unnecessarily requires the use of larger actuators resulting in higher power requirements。

Optimum robot design has been addressed by many researchers as found in the open literature based on kinematic and/or dynamic specifications。 Oral and Idler optimized a planar robotic arm for a particular motion。 They described the minimum weight design of high-speed robot using sequential quadratic programming [I]。 Yang and Ming-Chen applied the minimized degree of freedom concept in a modular robotic system in order to maximize the load carrying capacity and reduce the power requirements using an evolutionary algorithm (a variation of GA) [2]。 Paredis and Khosla discussed optimum kinematic design for serial link manipulators  using  simulated  annealing  based  on task specifications [3]。 Tang and Wang compared  Lagrangian  and  primal—dual neural network for real time joint torque optimization of kinematically redundant manipulators。 They used desired acceleration of end-effector as input for  a  specific task [4]。 Chen and Cheng described various control schemes and explained advantages of “minimum velocity norm method” over other methods for local joint torque optimi- zation in redundant manipulator [5]。 Paredis used a distributed agent based genetic algorithm (GA) to create fault tolerant serial chain manipulators  from  a  small inventory of links and modules [6]。 Chedmail and Ramstein used  a GA to determine the base position and type (one of several) of a manipulator to optimize workspace reachability [7]。

The major disadvantage of conventional optimization techniques is that they work from point to point, using local information to decide which  point  to explore  next。 This could lead to the premature convergence of optimization process at false optimum。 Also, many of the conventional search techniques require specific knowledge of the problem to be analyzed, for example gradient-based techniques required derivative information。 Another drawback of traditional optimization techniques is the require- ment for good initial guesses for the design  variables。