An Approach to Automatic Robot Programming Based on Inductive Learning

Abstract

Automatic programming of manipulator robots attracts in creasing interest within the context of mechanical assembly. In this paper we focus on the problem of mating two parts, which requires sensor-based strategies to deal with geometric uncertainty. We present a system that embodies a two-phase approach to build robot programs implementing such strate gies. First a training phase interacts with the robot actuatorsand sensors and produces traces of execution of a given part- mating operation. Next a purely computational induction phase transforms these traces into an executable manipula tor-level program for the operation. The system embedding this approach has been completely implemented, and it has been used experimentally on several assembly tasks.