Multimodal Object-Based Environment Representation for Assistive Robotics

摘要

Autonomous robots are nowadays successfully used in industrial environments, where tasks follow predetermined plans and the world is a known (and closed) set of objects. The context of social robotics brings new challenges to the robot. First of all, the world is no longer closed. New objects can be introduced at any time, and it is now impossible to build an exaustive list of them nor having a precomputed set of descriptors. Moreover, natural interactions with a human being don't follow any precomputed graph of sequences or grammar. To deal with the complexity of such an open world, a robot can no longer solely rely on its sensors data: a compact representation to comprehend its surrounding is needed. Our approach focuses on task independent environment representation where human-robot interactions are involved. We propose a global architecture bridging the gap between perception and semantic modalities through instances (physical realizations of semantic concepts). In this article, we describe a method for automatic generation of object-related ontology. Based on it, a practical formalization of the ill-defined notion of "context" is discussed. We then tackle human-robot interactions in our system through the description of user request processing. Finally, we illustrate the flow of our model on two showcases which demonstrate the validity of the approach.