Intelligent robot hand control system using a tailorable parallel computer concept

Presents a control system for the intelligent force control of multifingered robot grips. The multilevel system architecture combines both a fuzzy-based adaptation level and a neural-based one with a conventional PID-controller that allows autonomous performance of fine manipulations of an object. Two kinds of fuzzy controllers are presented. They use a decision making logic which expresses the a priori knowledge about the grasp force behaviour inside the friction cones and the necessary reactions regarding the criterion for grip stability. A neural controller, based on a Hooke-Jeeves optimisation approach, has been developed as well. The neural control algorithm is implemented by a three-layered backpropagation neural network. The neural and fuzzy controllers can be used separately or in parallel. In the last case the neural controller can be on-line trained using the input-output information from the fuzzy one. A computer based simulation system for the peg-in-hole insertion task is developed to analyse and to compare the capabilities of both control algorithms. The demands of flexibility and real-time control of the implementation of the control system are suited by a tailorable parallel computer concept. The two basic ideas of the concept are to set up each processing element individually for its application and connect these elements with different communicational methods according to the applicational demands. As this happens before runtime the concept is called static flexibility and is implemented using a new mechanical computer structure