TR-133: NARM: The Design of a Neural Robot Arm Controller

Abstract

This paper presents an approach for the collision-free comrol of general robot manipulators moving among a changing set of obstacles. The Neural Adaptive Robot Manipulator (NARM) controller, based on a layered, neural network architecture, adapts to the specific eye/hand and arm/body kinematics of any arbitrarily shaped robot during an initial, unsupervised training phase. After training, the robot selects a target point by “glancing” at it and the controller moves the end-effector into position. Collision-free movement is produced regardless of the number or arrangement of obstacles in the workspace. Moreover, no additional learning is required if the obstacle set is changed. This approach has several advantages over traditional algorithmic solutions and extends previous work on neural manipulator control. Results of a simulation are presented.