An experimental approach to robotic grasping using a connectionistarchitecture and generic grasping functions

摘要

An experimental approach to robotic grasping is presented. Thisnapproach is based on developing a generic representation of graspingnrules, which allows learning them from experiments between the objectnand the robot. A modular connectionist design arranged in subsumptionnlayers is used to provide a mapping between sensory inputs and robotnactions. Reinforcement feedback is used to select between differentngrasping rules and to reduce the number of failed experiments. This isnparticularly critical for applications in the personal service robotnenvironment. Simulated experiments on a 15-object database show that thensystem is capable of learning grasping rules for each object in a finitennumber of experiments as well as generalizing from experiments on onenobject to grasping from another