Neural network-based vision for precise control of a walking robot

Abstract

This article describes a connectionist vision system for the precisecontrol of a robot designed to walk on the exterior of the spacestation. The network learns to use video camera input to determinethe displacement of the robot's gripper relative to a hole in whichthe gripper must be inserted. Once trained, the network's outputis used to control the robot, with a resulting factor of five fewermissed gripper insertions than occur when the robot walks withoutsensor feedback. The neural network visual feedback techniques describedcould also be applied in domains such as manufacturing, where preciserobot positioning is required in an uncertain environment.