Novelty detection for semantic place categorization
André Susano Pinto · 2011
For a long time humanity has dreamed that one day robots will be among us. They will explore our world and interact with us, understand our concepts and reason. An important step in that direction is endowing robots with knowledge about human concepts and semantics. However, it is unrealistic to believe that the human world can be fully modeled in the robot’s brain at the design stage. Therefore, robots must be able to adapt and learn when confronted with novel situations. Detection of novel situations, where the knowledge of the robot is not sufficient plays an important role in adaptation and learning of new concepts and is the main topic of this thesis. In the context of mobile robotics, spatial concepts and semantics are crucial to enable the robot to perform complex human-like tasks and human interactions. For handling those, a robot builds a representation of space extended with semantic properties, process which is known as semantic mapping. This representation identifies spatial entities and classifies them according to their meaning to humans allowing the robot to reason at a high abstraction level. For example, humans categorize spaces as kitchens, bedrooms,