Learning inverse kinematic solutions of redundant manipulators using multiple internal models

TLDR: This work proposes to use multiple paired forward and inverse models approach described in the following sections, to obtain multiple IK solutions and points towards task space representation of motion in the brain.

Abstract: Biological systems are superior compared to robotic systems in their ability to adapt to new situations very quickly. Hence, it would be advantageous to take insights from the architecture of sensory-motor maps in designing controllers for robotic systems. Any movement can be represented either in task space or joint space of a given manipulator. Planning and control in task space essentially reduces the computational complexity compared to joint-space approaches due to fewer dimensions involved. Experimental evidences [1], point towards task space representation of motion in the brain. The transformation of these task space representations into joint space is however not trivial, as it forms an ill-posed problem. This constitutes the inverse kinematics (IK) problem for a given manipulator. We propose to use multiple paired forward and inverse models approach described in the following sections, to obtain multiple IK solutions.