Designing teachable robots

Abstract

This thesis advances the design of teachable adaptable robots. I propose two paths of improvement to the popular, easy to use, leading method, in which a teacher literally leads the robot by its hand through movements. The improvements enable motor commands to be changed and conditional branches to be formed, without the need for a keyboard or other explicit programming device. On improvement path 1, the addition of a verbal correcting (VC) scheme would enable a teacher to make on-line verbal corrections to a robot's movement sequences. The further addition of a production system of corrections (PSC) would enable a robot to remember and use verbally taught conditional corrections. On path 2 a goal-seeking (GS) system and VC would enable a teacher to set goals, lead movements, and verbally correct the robot. The robot then selects its own motor commands for achieving goals. A multiple context learning system (MCLS), a multiple, extended GS system, combines the two paths. It enables both sequences and goals to be taught to a led robot. A simple, but real, led MCLS-robot is demonstrated. I establish four important properties of MCLSs: (a) an MCLS can enable a robot to learn to perform motor commands that are initially performed only by reflex, so that eye and speech motor commands, neither suitable for being led, can still be learned; (b) an MCLS can learn to be a Turing machine, which is a universal computing machine, explicitly showing the error in criticisms of MCLSs' computational power; (c) the selections of a context learning system in an MCLS converge on the optimal motor commands for achieving goals; and (d) an MCLS-robot can handle the negation problem; doing something positive in the absence of a certain condition.