Unsupervised online learning for long-term autonomy. (December 2013)

• Record Type: Journal Article
• Authors: Ott, Lionel; Ramos, Fabio
• Is Part Of: International journal of robotics research. Volume 32:Number 14(2013)
• Page Start: 1724
• Page End: 1741
• Publication Date: 2013-12
• Subjects: Affinity propagation -- clustering -- long-term autonomy -- unsupervised learning; Robots -- Periodicals; Robots, Industrial -- Periodicals; 629.89205
• DOI: 10.1177/0278364913505657
• Languages: English
• ISSNs: 0278-3649

Abstract

A reliable representation of the environment a robot operates in is vital for solving complex tasks. Models that represent information about objects and their properties are typically trained beforehand using supervised methods. This requires intensive human labeling which makes it time-consuming and results in models that are generally inflexible to changes. We would prefer a robot that can build a model of the environment autonomously by learning the different objects and their corresponding properties without human supervision. This would enable the robot to adapt to changes in the environment as well as reduce the effort of deploying a robot to a new environment. In this paper we present solutions to these problems based on novel extensions of affinity propagation; a clustering method that can be executed in real time to produce meaningful models from observations gathered by a robot. Our method is applied to two different tasks. We demonstrate how to automatically learn models for predicting collisions from raw laser data. Then, the method is used to learn visual appearance models of the environment to recognize and avoid obstacles. In both cases, there is no human supervision; the methodology is entirely based on sensory information gathered by the robot and its interaction with the environment. In experiments we show how meta-point affinity propagation performs similarly to standard affinity propagation, while being faster and capable of handling much larger data-sets. Furthermore, we show how different features influence the prediction quality of the model for collision prediction from laser scans. Finally, we show how we successfully build and maintain an appearance model for obstacle detection which can be used to detect obstacles well before a collision could occur.