Organization and prioritization of sensory information using an ego-centered, locally-connected network

Abstract

A multimodal attentional system for a robot operating in a dynamic environment is described. The system at once enables a robot (a) to maintain a Focus of Attention (FOA) on the region of space most important to its current task, (b) to shift that focus to attend to new stimuli, and (c) to ignore repetitive stimuli of no importance nor of any danger. That is to say, the system exhibits both sensitization and habituation. The system is both task- and data-driven. It is a collection of abstract data types that form a locally-connected network of processing nodes and an FOA object that is connected to all. Each connection permits two-way data flow. The connectivity of the nodes forms a geodesic tessellation of a virtual sphere centered on the base frame of the robot. Parallel independent sensory processing modules (SPM) write data to the nodes as a function of the direction from which the sensory stimuli originated. Those data include abstracted sensory information, a measure of data's importance, the direction from which it was received, and the time of reception. Each node: (1) can receive data from any of the sensory processors and of any modality, (2) has a two-stage memory for each modality, (3) computes, modulates, and maintains an activation value for each modality as a function of the data and its importance, and (4) communicates with its neighbors to share sensory data and activation values. The FOA maintains a prioritized list of tasks and an identifiers for the node with the largest active and the node with the largest current change in activation. The aggregate of simple local processes at each node together with a global task specification cause an FOA to emerge with the properties (a)-(c) above. Background on the network, on attention, and on habituation are presented. Experiments are described that demonstrate the strengths and weaknesses of the approach.