Swarm-Based Active Vision

This paper proposes a computational distributed model for active vision. In the proposed model, action selection and visual processes progressively unfold in a parallel and asynchronous way through a set of cross-modulatory signals. The visual process is modelled with a swarm of perceptual agents inhabiting the physical agent’s sensorimotor space, motivated by the ant foraging metaphor. Perceptual agents, called perceptual-ants (p-ants), perform local active vision, whereas the self-organised collective behaviour maintains global spatio-temporal coherence, i.e. a social cognitive map. A by-product of the method is the ability to maintain distributed, active and sparse spatial working memories, i.e. local maps of the environment. Experimental results with a simulated robot, performing a simple navigation task, show the ability of the model to perform both robustly and parsimoniously in terms of processing.