Grounding Language to Natural Human-Robot Interaction in Robot Navigation Tasks

Abstract

With the development of service robot technology, the natural interaction between humans and service robots has attracted increasing attention. People are no longer satisfied with using buttons or controllers to issue instructions and using natural language to guide the navigation of service robots has become a hot research topic. At present, most methods rely on corpora or labeled datasets with a large amount of data, which usually require huge labor costs. To address this problem, this paper proposes a robot navigation method based on natural language interaction in known environments. In the robot navigation instructions, the tasks are usually determined by the landmarks and the temporal words. In this work, the landmarks of a known environment is firstly extracted and a retrieval table of these landmarks is constructed, which is utilized to further landmark extraction from the natural language instruction. Then the landmarks and the temporal words are extracted from the instruction, sort the temporal words and their corresponding landmarks by their weights. Finally, the sorted landmarks are then published as navigation goals one by one and guide the robot in navigation. The experimental results show that our method is capable of guiding the robot navigation with natural language instructions in a known environment.