抄録
This paper presents a new approach to computing multiple inverse kinematic solutions for redundant manipulators by inverting modular neural networks. This approach is a three-phase procedure. In the first phase, the configuration space is partitioned into a set of regions to learn the forward kinematic function of a manipulator, and a set of modular neural networks is trained on associated sets of input-output data sampled over these regions. In the second phase, multiple inverse kinematic solutions for a desired end-effector position are computed by inverting the corresponding trained modular neural networks. In the third phase, an “optimal” solution is selected from the multiple inverse kinematic solutions according to a given criterion. The important advantage of this approach over existing methods is that an “optimal” inverse kinematic solution can be obtained from the multiple solutions. Therefore, better control of the manipulator can be achieved. This approach is illustrated with a three-joint planar arm.
