A robot grasp synthesis tool based on neural computational models

The paper deals with the grasp synthesis problem arising in task planning for robotic dextrous hands. To this end, a neural architecture is proposed to exploit sources of knowledge not easy to formalize, such as practical experience learned through experiments with a real device. After a preliminary discussion of representation levels and techniques, the paper describes the design issues of a neural tool for selecting the feasible grasps of a robotic hand in various situations and ranking them according to object geometric criteria. The neural tool takes advantage of the interaction with a rule-based expert system, which exploits heterogeneous sources of knowledge such as heuristic rules, geometric reasoning and, when applicable, analytical results, in order to determine the object shape and size. Moreover, the expert system takes into account the workspace constraints and task requirements to select the object surface pairs candidates to pose the hand contacts. The validity of the proposal has been ascertained with reference to the University of Bologna hand, but the underlying methodological issues can be specialized to other robotic dextrous hands