Visual Semantic Robot Navigation in Indoor Environments

Abstract

This work proposes a robot navigation framework based on semantic information extracted from visual data. The robot rather than navigating by following a sequence of metric poses, is able to reach its targets by activating behaviors described by natural language (e.g. get out of the room, follow the corridor, and then enter the room at the right). The system operates without prior map knowledge or a metric representation. This type of navigation is inspired from humans, where places are not described in terms of a global map but with semantic information. One of the advantages of this representation is that the system is able to recover from errors occurring in case of conflicts with the current context. E.g. a corridor-centering behavior activated inside a room.