DEMONSTRATION INSTEAD OF PROGRAMMING: FOCUSSING ATTENTION IN ROBOT TASK ACQUISITION

Abstract

This thesis is an important advance in making robots more useable. It develops a task acquisition system which demonstrates the feasibility of constructing new programs just from the user leading a robot. The result is ETAR, for Example-based Task Acquisition in Robots, and has been implemented on an Excalibur robot. Any person, who can do a task with the common direct lead mechanism on industrial manipulators, can designate it to the robot through ETAR. Thus, ETAR is an alternative to robot programming. The acquired procedures are not only repeated sequences, but standard assembly tasks such as widget construction and block stacking--tasks with loops, branches, and variables. ETAR is a prototypical machine learning system which begins from user examples on a real robot, requires minimal background knowledge, learns inductively, and generates the task description with the aid of a focussing mechanism. The focussing mechanism forces ETAR to concentrate on important domain objects, thus eliminating useless steps, determining a symbolic translation for the task, finding loops, introducing branches, and inducing functions to merge examples into one general program. Additionally, this thesis contributes to low level robotics. It provides unpublished kinematics for the Excalibur robot. Furthermore, it offers a unique, intuitive introduction to quaternions and describes how they rotate vectors and interpolate orientations more efficiently than matrices. Quaternions are used to obtain straight line motion for the Excalibur robot. Implementing kinematics and motion interpolation was a preliminary requirement to the learning algorithm.