Learning of inverse kinematics behavior of redundant robot

Abstract

Acceleration in redundancy resolution has been achieved by model-based learning. Parameterized joint motions generated by pseudoinverse of parameterized motion primitives in operational space are taken as learning examples. The method of successive approximations (SA) as a function approximator generalizes given examples resulting in a parameterized model, termed redundancy resolution skill. The SA method results in an analytical solution of joint motions. This particular property enables extension of the universality of skills to tasks never experienced by learning. Moreover, a proper addressing of skills output results in paths completely different from learned ones. In practice, this means that a segment as a motion primitive suffices in acquiring skill in the learning domain. The example emphasises this property of SA-based redundancy resolution. Highly compressed skill, acquired on segments only, is applied in RR along two test-trajectories-an 'H'-character and a rosette. The gained acceleration of RR recommends this procedure for online redundant robot control and easy robot programming.