Efficient Learning of Generic Grasping Functions using a Set of Local Experts

In this article, we presented the concept of generic grasping functions that can be used to find grasping choices for arbitrary objects in the environment Examples of these functions actions can be collected from sensory data and approximated using artificial neural networks. We tested the application of a modular neural architecture to the approximation of some simulated grasping functions. A modified architecture in which the gating network output rule was replaced by a competitive output rule gave better results. We also found that applying an incremental learning process would significantly enhance the learning rate and the approximation accuracy.