Regularization of inverse kinematics for redundant manipulators using neural network inversions

This paper presents a new approach to regularizing the inverse kinematics problem for redundant manipulators using neural network inversions. This approach is a four-phase procedure. In the first phase, the configuration space and associated workspace are partitioned into a set of regions. In the second phase, a set of modular neural networks is trained on associated training data sets sampled over these regions to learn the forward kinematic function. In the third phase, the multiple inverse kinematic solutions for a desired end-effector position are obtained by inverting the corresponding modular neural networks. In the fourth phase, an “optimal” inverse kinematic solution is selected from the multiple solutions according to a given criterion. This approach has an important feature in comparison with existing methods, that is, both the inverse kinematic solutions located in the multiple solution branches and the ones that belong to the same solution branch can be found. As a result, better control of the manipulator using the optimum solution than that using an ordinary solution can be achieved. This approach is illustrated with a three-joint planar arm