Semantic place labeling with mobile robots

Abstract

Indoor environments can typically be divided into places with di?erent functionalities like corridors, rooms or doorways. The ability to learn such semantic categories from sensor data enables a mobile robot to extend the representation ofthe environment, and to improve its capabilites. As an example, natural languageterms like corridor or room can be used to communicate the position of the robotin a more intuitive way. Other tasks, like exploration or localization, can also becarried out by the robot in a better way when semantic information is taken intoaccount.In this thesis, we present a method that enables a mobile robot to classify thedi?erent places of indoor environments into semantic classes, and then use this information to extend its representations of the environments. The main idea is toclassify the position of the robot based on the current observations taken by therobot. In this work, we use as main observations the scans obtained from a laserrange sensor. Each scan is represented by a set of features that encode the geometrical properties of the current position. These features are then used to classify thescan into the corresponding semantic class. The output of the classi?cation is represented by a probability distribution over the set of possible semantic classes. Thisprobabilistic representation permits us to apply further probabilistic techniques toimprove the ?nal classi?cation, reducing the number of errors. We also presentan extension which enables the robot to include other types of observations in theclassi?cation, like camera images.This work additionally introduces several applications of the previous approachin di?erent robotic tasks. First, we will show how the semantic information can beused to extract topological maps from indoor environments. In a second application, we present a method that incorporates transitions between di?erent placeswhen classifying a trajectory taken by a mobile robot. It will also be shown that thesemantic information can reduce the time needed by the robot in exploration andlocalization tasks. Finally, we present the semantic classi?cation of places as partof an integrated robotic system designed for interacting with humans using naturallanguage.