David DeMers, Kenneth Kreutz-Delgado
We introduce and demonstrate a bootstrap method for construction of an in(cid:173) verse function for the robot kinematic mapping using only sample configuration(cid:173) space/workspace data. Unsupervised learning (clustering) techniques are used on pre-image neighborhoods in order to learn to partition the configuration space into subsets over which the kinematic mapping is invertible. Supervised leam(cid:173) ing is then used separately on each of the partitions to approximate the inverse function. The ill-posed inverse kinematics function is thereby regularized, and a globa1 inverse kinematics solution for the wristless Puma manipulator is devel(cid:173) oped.