A Robot Command Learning System Acorn-II and an Evaluation of This System

Abstract

There has been research for intelligent robots that has focused on improving robot's performance and making robots execute several tasks reliably. This research is important ; however, the research for human-robot intelligent interaction will also be needed for the application of intelligent robots in the future. People expect to communicate with intelligent robots via natural language. However, because the mearning of many abstract symbolic representations changes according to the situation, it is difficult to make a robot understand the representation and execute several tasks according to the situation. This paper proposes Acorn-II, a system that learns relational behavior and situations. The system incrementally learns algebraic expressions that represent the constraints among sensors and actuators from positive and negative instances of sensor/actuator data. Information for recognizing external situations can not be perceived directly from primitive features. It can be perceived from new features constructed from primitive features. So, this system learns algebraic expressions that represent the constraints among primitive features by combining a "Feature Construction" method and a "Generalization To Interval" method. The paper evaluates this system with empirical results and shows that it can learn robot commands more accurately than the famous learning systems CN2 and C4. It has been implemented in an autonomous mobile robot Einstein I.