PSOM network: Learning with few examples

Precise sensorimotor mappings between various motor, joint, sensor, and abstract physical spaces are the basis for many robotics tasks. Their cheap construction is a challenge for adaptive and learning methods. However, the practical application of many neural networks suffer from the need of large amounts of training data, which makes the learning phase a costly operation - sometimes beyond reasonable bounds of cost and effort. In this paper we discuss the "Parameterized Self-organizing Maps" (PSOM) as a learning method for rapidly creating high-dimensional, continuous mappings. By making use of available topological information the PSOM shows excellent generalization capabilities from a small set of training data. Unlike most other existing approaches that are limited to the representation of a input-output mappings, the PSOM provides as an important generalization a flexibly usable, continuous associate memory. This allows to represent several related mappings - coexisting in a single and coherent framework. Task specifications for redundant manipulators often leave the problem of picking one action from a subspace of possible alternatives. The PSOM approach offers a flexible and compact form to select from various constraint and target functions previously associated. We present application results for learning several kinematic relations of a hydraulic robot finger in a single PSOM module. Based on only 27 data points, the PSOM learns the inverse kinematic with a mean positioning accuracy of 1 % of the entire workspace. Another PSOM learns various ways to resolve the redundancy problem for positioning a 4 DOF manipulator.