Aspects of Visual Attention for Autonomous Mobile Robots

Abstract

The deployment of technical systems in complex and unstructured everyday environments has become an essential direction of robotic research, where the limited computation capacity and the real-time requirements become the bottleneck of the system development. Cognitive abilities to interpret and select essential information from a large amount of sensory data are important and necessary, especially for a mobile robotic system. From the extensive works in biology, cognitive psychology, and neuroscience, visual attention is considered to be one of the most powerful cognitive processes dealing with visual information selection. Considering the challenges arising in the aforementioned context, both biologically plausible and technically applicable robot visual attention strategies should be developed to bridge the gap between fundamental studies and specific technical realizations. This thesis focuses on the investigation of goal-directed visual attention strategies for autonomous mobile robots, explored from three different perspectives: the stimulus-dependent aspect, the task-relevant spatial aspect, and the task-relevant temporal aspect. Two information-based metrics are proposed to enable well-timed perception of temporal and spatial stimuli, which is a critical factor for awareness of unexpected events and for ensuring the working order of robots. Integrated approaches to top-down and bottom-up attention selection are elaborated, where the determination of robot spatial attention allocation for task-relevant information is investigated. A human-inspired temporal attention control strategy is proposed, considering the challenge of a limited field of view in multi-object tasks. Evaluation and demonstration are carried out in simulations and experiments. The main contributions are qualitative improvements of sensitive awareness of environment dynamics, efficient, flexible, and adaptable enhancement of task-relevant information, as well as significant reduction of the overall perception uncertainty through temporal attention planning. In this work, application-oriented attention control considering characteristics of mobile robots functioning in dynamic environments is studied in a general and integrated manner for the first time. The contributions advance the state of the art in cognitive robot design and provide valuable insights for future research.