Inverse Kinematic Solution in Handling 3R Manipulator via Real-Time Genetic Algorithm

F.Y.C. Albert 1, S.P. Koh1, S.K. Tiong1, C.P. Chen1, F.W. Yap2 1 Department of Electronic & Communication Engineering, Universiti Tenaga Nasional, Km 7, Jalan Kajang-Puchong, 43009 Kajang, Selangor. albert@uniten.edu.my, johnnykoh@uniten.edu.my, siehkiong@uniten.edu.my, chaiphing@uniten.edu.my, 2 Faculty of Engineering and Technology, Multimedia University, Jalan Ayer Keroh Lama, Bukit Beruang, 75450 Melaka, Malaysia. fwyap@mmu.edu.my Abstract
The Artificial Intelligence (AI) machine usage examines various benefits of its applications in industries. This research is to develop an AI 3 revolute (3R) planar manipulator system that would contribute to a productive and a quality way of material handling and processing. The system would adopt the advantage of Genetic Algorithm to optimize its performance in terms of motion control and accuracy. The GA would use the new proposed crossover operator, namely Dynamic Multilayered Chromosome Crossover (DMCC). The paper presents simulation results obtained based on the information of the targets position and the manipulator orientation. principles of evolutionary biology applied to computer science. The kinematics is the process of solving and calculating the position and orientation of the endeffector. Basically, when given the end point of the end-effector, the angles of the joints are determined to achieve the required destination. It can be difficult, and it would usually have many or infinitely many solutions. Moreover, it involves time-consuming calculations to solve the inverse kinematics problem by the numerical iterative method if the dimension of the Jacobian matrix becomes large. From the literature, there is much research being pursued to solve the path planning problem. A significant amount of research also has been reported regarding to the trajectory planning for redundant degree of motion freedom robot manipulators [1,2]. However, most of them are based on the analytical methods of inverse kinematics employing a pseudoinverse of the Jacobian matrix. Doyle and Jones [3] proposed a path-planning scheme that uses GA to search the manipulator configuration space for the optimum path. The result indicated that the GA generates good path solutions. Davidor [4] applied genetic algorithm (GA) to the trajectory generation by searching the inverse kinematics solutions to predefined end-effector robot paths. Pires and Machado [5] proposed an evolutionary method which optimizes the robot structure and the required manipulating trajectories. They described how an optimal manipulator minimizes both the path trajectory length and the ripple in the time evolution, without any collision with the obstacles in the workspace. An algorithm containing a genetic algorithm and a pattern

1. Introduction
In the study of robotics and automation in the industry environment, many researchers have worked to develop theories and algorithms for the important path planning problems in the manufacturing automation field. There are number of algorithms, strategies and their variations that are currently being used for solving complex optimization problems. So, a proposed form of genetic algorithm started to be applied in the systems in order to optimize the desired objectives. Genetic Algorithm (GA) represents one of the artificial intelligence algorithms that are used to improve performance in the systems. GA is a technique for efficiently finding near optimal or approximate solutions relating to optimization and difficult problems through a set of techniques operation of selection, crossover, and mutation. Such techniques are

978-1-4244-2328-6/08/$25.00 2008 IEEE

search is introduced to design the optimal point-topoint trajectory planning for a planar 3-DOF manipulator. Genetic algorithms have been shown to be effective in solving NP-hard problems, thus they are often used for path planning [6]. This paper work focuses on handling and optimizing the inverse kinematics for real-time trajectory planning manipulator based on genetic algorithm. A planar of three degree of freedom manipulator system has been chosen for this study in order to test the reliability of the proposed GA.

2. Problem Formulation
Given the initial and target position that is defined by the user. The problem is to find the optimum end tip hand point value and the optimum joint angles as a function of the goal (wrist position x, y and orientation ) for the robot manipulator to reach the final positions. Let {1, 2, 3} be the joint angles of a 3-link serial manipulator with link lengths {L1, L2, L3} as shown in Figure 1. All the joints are allowed to rotate at a maximum angle of 360.

Figure 2: Flowchart for Inverse Kinematics Genetic Algorithm (IKGA). The first step is to encode the unknown joints angle, represented in binary as strings of 0s and 1s [7]. After the encoding process, the GA would proceed to initialize a population of solutions randomly, and then improve it through repetitive application of selection, crossover and mutation operators. The GAs uses a stochastic sampling which is also called the roulette wheel selection method. This sampling method selects parents from the current population according to a spin of a weighted roulette wheel. The fitness function is formulated based on the difference from target object to the end-effector. The value of the fitness function would exponentially reduce if the difference is small. The fitness of the solution is expressed as in equation (3.1).
fitness = 1 exp(

(X tcp X obj )2 + (Ytcp Yobj )2 ) (3.1)

Figure 1: The joint variables and link lengths for a 3R planar manipulator.

3. Methodology
In this paper, the synthesis of inverse kinematics and genetic algorithm would be implemented. The approach method for the inverse kinematics genetic algorithm is depicted in Figure 2.

Where the point (Xtcp , Ytcp) is the manipulator endeffector point, while the point (Xobj , Yobj) is the object target position in the workspace. The is used as the control parameter to fine tune the efficiency of the fitness equation. The parents were selected based on their fitness, where whoever with greater fitness would be most likely to be selected for breeding. The Dynamic Multi-Layered Chromosome Crossover (DMCC) is used in the crossover operation to produce two offspring from the selected two parents. The new proposed crossover operator would be dynamically selecting for two points and would determine the layers which to be applied with the crossover operation. Every gene that between the two points would be swapped among the parents strings, to create two new offspring. The DMCC operation process is illustrated

978-1-4244-2328-6/08/$25.00 2008 IEEE

in Figure 3. For mutation operation, a random gene would be selected to be mutated. The mutation would choose the best string or individual from the populations (parents). This process would change the value of the selected gene from the individual string.

j =

c 2n 255 n =0

180 o

(3.2)

4. Simulation
A simulator package has been developed to evaluate the 3R planar manipulator IKGA, as depicted in Figure 3. It would explore and search the object target in the workspace and try to obtain the unknown orientation of the manipulator. The simulation has been carried out using the GAs parameters as given in Table I.
Table I: GA simulation parameter Simulation Parameter Value Maximum Generation 100 Population, po 50 Chromosome length/layer 9/3 Selection Method Stochastic Crossover Rate, pc 80% Mutation Rate, pm 0.025 Mutation Point, mp 2 No. of Best Chromosomes Kept, kb 1 Crossover Type DMCC

Figure 3: DMCC operator operation. The fitness of the new breeding generation after the crossover and mutation operations would be evaluated. Consequently, the current good ranked individual would be replaced with the best fitness offspring. This generational process would be done iteratively until the termination condition is met. The process would be terminated if a solution is found that satisfies the minimum criteria that can be either the minimum end point accuracy or the maximum number of generations assigned. Eventually, the best solution would then be decoded using equation (3.2). In this equation, the j is the joints angle value and the variable c can be either 0 or 1 dynamically. The first binary bit from the chromosome string represented by the mathematical sign () in the equation would determine the rotating direction of the joints angle.

The manipulator base is placed at the origin point of (Xo=3500mm, Yo=3500mm). The default base point is at the center of the workspace. In the simulation, it would be impossible for the linked structure or the endeffector to touch the object if the located target object is out the working envelope. If the target object is within the workspace range, there would be one or two solutions for the two-link manipulator system. But if there are more than two joints, there would frequently be infinitely many solutions to the problem. The links would have a small amount of springiness; this shows that the links keep moving even when it has found the object target. In other words, it is trying to get into a more comfortable and appropriate angle position.

978-1-4244-2328-6/08/$25.00 2008 IEEE

Figure 4: GUI simulation for 3R planar manipulator IKGA.

The free body diagram (FBD) of the 3R planar manipulator obtained from the simulator is depicted in Figure 5. Figure 5(a) illustrates the link structures in the home position. The following Figures 5(b), 5(c) and 5(d) illustrate the manipulator in trying to reach the targets represented by the small square dots {O1, O2, O3} in the workspace.

5. Preliminary Results
From the simulation observation, the end-effector managed to reach the target with the angle result obtained as given in Table II.
Table II: Simulated Manipulator Angle Angle Parameter O1 O2 O3 1 0 43.24 40.16 2 48.71 117.15 119.86 3 79.47 36.59 119.80

Figure 5: FBD of 3R planar manipulator reaching the targets.

Table III shows the average of the fitness value and the number of generations required tested with different types of crossover operations. Besides DMCC, the system has been tested using Dynamic Point Crossover (DPC) and Random means. In DPC, the crossing point and the size of the crossover region are dynamically chosen, but with a fixed crossover layers. All the layers would be selected for crossover in DPC. The results show that the implementation of DMCC operator has outperformed the others.

978-1-4244-2328-6/08/$25.00 2008 IEEE

Table III: Effects of Different Crossover

Operations Type DMCC DPC Random Generations Required 25 47 63 Fitness 0.0041 0.0122 0.0131

Figure 6: Result of DMCC operation.

978-1-4244-2328-6/08/$25.00 2008 IEEE

7. Acknowledgement
This work was performed with the support of the MOSTI (Ministry of Science, Technology and Innovation, Malaysia).

8. References
[1] D. E. Goldberg, Genetic Algorithms, New York: Addison-Wesley, 1989. [2] J.Y.S. Luh, Y. L. Gu, Industrial Robots with seven Joints, Proc. IEEE International Conference on Robotics and Automation, Mar. 1985.

Figure 7: Difference values obtained Vs number of generations.

[3] A.B. Doyle, D.I Jones, Robot Path Planning with Genetic Algorithms, Proc. 2nd Portuguese Conf. on Automatic Control, pp. 312-218, Porto, Portugal, 1996. [4] Y.Davidor, Genetic Algorithms and Robotics, a Heuristic Strategy for Optimization, World Scientific Publishing, 1991. [5] E. J. Solteiro Pires, J. A Tenreiro Machado, P.B. de Mour Oliveira, An Evolutionary Approach to Robot Structure and Trajectory Optimization, Proc. 10th International Conference on Advanced Robotics, Budapest, Hungary, pp. 333-338, 22-25/Aug/2001. [6] D. A. Ashlock, T. W. Manikas, K. Ashenayi, Evolving a Diverse Collection Robot Path Planning Problems, Proc. IEEE Congress on Evolutionary Computation, Vancouver, BC, Canada July 16-21, 2006. [7] D. A. Coley, An Introduction to Genetic Algorithms for Scientists and Engineers, World Scientific Publishing, 1999. [8] L.Davis, Handbook of Genetic Algorithms, New York, Van Nostrand Reinhold, 1991.

As can be observed in Figure 7, shorter time is required for DMCC over the others to achieve the convergence. At some earlier points, better individuals were eventually born, and the algorithm was resuming the improvement process. The number of generations required for solving the problem is 25 for DMCC, 47 for DPC and 63 for Random. In addition, the solution for DMCC is more accurate compare to the others, achieving a difference of 0.0041. Figure 7 shows the result of the DMCC operation. The multi-layered parents and its offspring are depicted in Figures 6(a) and 6(b), respectively. In DMCC, the crossover operator would dynamically selecting two points from each of the layer (1, 2, 3) and determines the layer(s) that need to be crossover. All the genes that are in between the two points would be swapped among the parents strings, to create two new offspring.

6. Conclusion
The proposed algorithm is useful when dealing with complex path planning in the automation systems such as for robotics in welding and laser scanning control applications. In this research, a simulator package has been developed to test the performance of the 3R planar industrial manipulators. Besides the new proposed DMCC, the system has been tested using DPC and Random means that are typically used in real time conventional GA. The results indicate that the problem of kinematics is solved for the system. The proposed method exhibits improved search speed, good accuracy and approximate solution.

978-1-4244-2328-6/08/$25.00 2008 IEEE