Dialogue-Based Supervision and Explanation of Robot Spatial Beliefs: a Software Architecture Perspective

摘要

The paper presents a software architecture allowing a robot to learn new compositions of objects in table-top scenarios by human demonstrations. The robot qualitatively represents those scenes, reason upon their similarity, and interact with humans through dialogues to talk about represented scenes. We formalise the robot behaviour based on a Description Logic representation of scenes through spatial beliefs, i.e., learned logic predicates, on which the robot applies symbolic reasoning to recognise and explain the scene. We exploit the logical structure of predicates in a software architecture that enables a robot exposing its beliefs, and if required, it allows a human supervisor to apply corrections in a form akin to robot active perception. The paper critically discusses the design of the software components and their interfaces, discriminating between knowledge representation and dialogue management. Those components are developed for human-robot knowledge sharing applications involving visual, verbal, and auditory modalities of interaction. Software components are treated as grey boxes managing an ontology-based formalisation of robot beliefs through four contextualised dialogues, for which we present a unique design pattern.