An Attention-Based Framework for Context Identification in Autonomous Robots

Being able to correctly assess the context it is currently acting in is a very important ability for every autonomous robot performing a task in a real world scenario such as navigating, manipulating an object or interacting with a user. Sensors are the primary interface with the external world and the means through which contextual knowledge is generated. Humans and animals use cognitive processes such as attention to selectively process perceived task-relevant information and to recognize the context they are currently acting in. Biologically inspired computational models of attention have been developed in recent years to be used as interpretation keys of mainly visual sensor data. This paper presents a new framework for situation assessment that expands existing computational models of attention by providing a unified methodology to interpret and combine data from different sources. The method utilizes probabilistic state estimation techniques such as Bayesian recursive estimation, Kalman filter, and hidden Markov models to interpret features extracted from sensor data and formulate hypotheses about different aspects of the task the robot is performing or of the environment it is currently acting in. The concept of Bayesian surprise is also used to mark the information content of each new hypothesis. A weight that takes into account the confidence in the estimate that generated the hypothesis, its information content, and the quality of the data is then calculated. The methodology presented in this paper is general and allows to consistently apply the framework to data from different types of sensors and to then combine their hypotheses. Once formulated, hypotheses can then be used for context-based reasoning and plan adaptation. The framework was implemented on a small two-wheel differential drive robot equipped with a camera, an ultrasonic and two infrared range sensors. Three different sets of results that evaluate the performance of different features of the framework are presented. First, the method has been applied to detect a target object and to distinguish it from similar objects. Second, the hypotheses strength calculation method has been characterized by isolating the effect of belief, surprise, and of the quality of the data. Third, the combination of hypotheses from different modules has been evaluated in the context of environment classification.