reMap: Spatially-Grounded and Queryable Semantics for Interactive Robots

Abstract

The semantic information available to a robot enhances its understanding of the world and allows it to adapt its behaviour accordingly. This information can be spatially-grounded, meaning it is associated with specific areas of the environment. For a robot to use this information effectively during its tasks, it is crucial to provide an efficient system for storing and retrieving spatially-grounded semantics. In this paper, we present reMap, a novel framework for the efficient representation, storage, and retrieval of spatially-grounded semantics. In reMap, we formally introduce Representation Maps (RMs), three-dimensional functions that each represent a different type of semantic information in space. These structures can be combined through operators to extract additional spatially grounded semantic information. reMap includes a SPARQL-based language that serves as a programmatic interface for retrieving spatially-grounded semantics stored in RMs. We provide an open-source ROS-based implementation of reMap, enabling efficient three-dimensional information storage and processing using dense voxel maps based on the high-performance OpenVDB format. Finally, we describe the execution of the framework over real-world data recorded in a semantically rich real-world environment.