An Attention-based Methodology for Context Identification and Exploitation in Autonomous Robot

Every intelligent agent, human or artificial, heavily relies on its understanding of the current context, that is on its situational awareness, to choose its next set of actions. Motivated by the often harmful outcomes of decisions made without appropriate situational awareness, a number of studies have been conducted in the field of ergonomics to understand how humans construct and utilize situational awareness when performing critical tasks. These findings constitute a solid base upon which computational methods that address the same problem in autonomous robots can be designed. In particular, when considering the issue in autonomous systems, three aspects need to be addressed and combined: perception, context classification, and context-dependent decision making. In fact, perception plays a major role in providing a robot with information to understand its surroundings and this makes the development of appropriate and efficient methods to analyze perceptual information of primary importance. Facing the same issue, primates have developed attention-based cognitive processes and a number of computational models have been designed to mimic this behavior in artificial agents. The development of appropriate methods for combining different types of often uncertain information into a coherent high level representation of the current context is also a very critical issue for the successful selection of appropriate actions. It is in fact this high level classification that allows an intelligent agent to reason about which goals to pursue or about how to adapt its behavior. Despite the multitude of context identification and exploitation methodologies proposed in literature, this is still an open research issue, lacking a comprehensive approach that considers all the components of the problem.;This dissertation aims at addressing the problem of developing a comprehensive methodology for context identification and exploitation in autonomous robots by using an attention-based approach. In particular it investigates how the concept of attention, traditionally used for salience detection when processing visual data, can be applied at different levels of a robotic architecture. To achieve this goal, first this work presents a framework for attention-based extraction of high level hypotheses from sensor data. Then, a method is presented that extends the concept of attention beyond processing of sensor data to the process of inferring the current context the robot is acting in. Last, a number of methods are presented that show how the acquired information can be used by an autonomous robot for goal selection and dynamic behavior adaptation. The attention-based framework for formulation of high-level hypotheses utilizes Bayesian inference to determine the likelihood of different states based on available information. Then, the concept of Bayesian surprise is utilized to quantify the novelty introduced by the latest acquired information. The current context is inferred by using the hypotheses as input evidence into a Bayesian network and then performing a cost-benefit analysis, modeled as Top-Down attention, on the network query variables. An architecture was developed to evaluate the interaction of the methods and their implementation on a robotic platform. The main drive when designing the architecture has been modularity in order to simplify its adaptation to different types of attention models and tasks. All the methods have been implemented in C++ and executed on the CPSBot, an educational two-wheel differential-drive robot controlled by an Arduino and a Raspberry Pi. Two case studies have been defined to evaluate the methods: robot soccer and surveillance. Future work will involve expanding the application of the method to teams composed of multiple robots, the use of machine learning to make the system more adaptable to new environments, and the application to autonomous vehicles and to tasks that require the interaction with a human user.