Indoor autonomous robot navigation by using visually grounded semantic instructions

Abstract

Navigation is becoming an increasingly substantial part of autonomy in robotics. As the capability of robots is continually researched and developed towards increasing complex tasks, navigation has become a topic that many are diving into. There has been numerous developments in navigation, from using Simultaneous Localization and Mapping to build maps in navigation, to exploring other forms of maps like Topological, Semantics and even Abstract maps. This project will work towards exploring a different kind of navigation that is essentially map-less. This is especially useful in foreign environments that the robot have no knowledge on whatsoever.

More specifically, this project focuses on defining the navigation goal as a mere string, of either a room name or a unit number. Without a map as a reference, the robot is forced to rely on extracting text information from its surrounding to identify its goal. Although the focus of this project is on goal-oriented navigation using text from the environment, the idea being explored in this project could be applied to many different functions and even used together with other existing work. This study is similar to previous work on extracting directional instructions for the robot to follow by extracting information from signs with the exception that it focus on finding the goal by verifying different potential end points, which are doors in this case, in the environment.