Optimal information distribution and performance in neighbourhood-conserving maps for robot control

A novel programming paradigm for the control of a robot manipulator by learning the mapping between the Cartesian space and the joint space (inverse kinematic) is discussed. It is based on a neural network model of optimal mappings between two high-dimensional spaces introduced by T. Kohonen (1982). The author describes the approach and presents the optimal mapping, based on the principle of maximal information gain. Furthermore, the principal control error made by the learned mapping is evaluated for the example of the PUMA robot. By introducing an optimization principle for the distribution of information in the neural network, the optimal system parameters, including the number of neurons and the optimal position encoding resolutions, are derived.<>