Learning by an autonomous agent in the pushing domain

Abstract

Zrimec, T. and Mowforth, P., Learning by an autonomous agent in the pushing domain, Robotics and Autonomous Systems, 8 (1991) 19–29. The work presented in this paper is concerned with developing an algorithm for the extraction and representation of knowledge which allows a real robot to learn predictable behaviour. We have designed an experiment in which a robot can randomly explore the domain of object pushing using signals recorded before and after a controlled movement. The goal is to let the system discover how its world domain works through experimentation and unsupervised learning. The learning method involves a combination of learning from examples along with partitioning, constructive induction and determination of dependencies which together allow the system to be independent of human help. Each experiment may be considered as a state transformation recorded as a sequence of attribute values. Treating each transformation as a training example, a large set of data was collected and subjected to the learning system. Robust and useful transformations were discovered which were represented as a hierarchical qualitative model. One important result is that representation in an actor-oriented, coordinate frame provides the most compact description for the problem domain. Unlike other domains, this style of experimentation offers the potential for closed-loop learning where, as far as the robot is concerned, its world is the oracle.