ABSTRACT
This chapter shows that existing KSOM-based visual motor control algorithms are inefficient for applications in redundant manipulators. The existing learning architectures do not preserve topology of the output space. Thus such algorithms become sensitive to initial network parameters. Since existing learning architectures do not preserve redundancy, the redundant manipulator can not perform dexterous tasks using these VMC algorithms. Thus a KSOM based redundancy preserving network proposed in the previous chapter is used to provide several kinematic configurations for a given target position. A real-time algorithm to learn network parameters has been proposed. Since each lattice neuron is associated with multiple solutions in joint angle space, an online adaptive clustering algorithm has been proposed to learn these joint angle vectors. It is shown that the proposed KSOM network is insensitive to initial network parameters unlike the standard KSOM network. The smoothness of joint angle trajectories is maintained through a modified neighborhood concept thereby preserving the conservative property of the inverse kinematic solution. We have also discussed about the all the improvements in the system compared with the existing mehods. Three criteria namely, lazy-arm movement, minimum angle-norm movement and minimum condition number of Jacobian matrices, are used to resolve redundancy. Finally, simulation results are validated through experiments on a 7 DOF PowerCube TM robot manipulator.
