| Resumo: | This thesis addresses the problem of modelling visual attention in the context of autonomous off-road robots. The goal of using visual attention mechanisms is to allow robots to focus perception on the aspects of the environment that are more relevant to the task at hand. As this work will show, this capability is a promoter of robustness and computational parsimony in both obstacle and trail detection. These features are key enablers of fast and energetically efficient field robots. One of the major challenges in modelling visual attention emerges from the need to ensure that the model is able to manage the speed-accuracy trade-off in the face of context and task changes. This thesis shows that this trade-off is handled if the cognitive process of visual attention is modelled as a self-organising process, whose operation is modulated by the robot’s action selection process. By closing the loop from the action selection process to the perceptual one, the latter is able to perform on a by-need basis, anticipating actual robot moves. To endow visual attention with self-organising properties, this work gets inspiration from Nature. Concretely, the mechanisms underlying the ability that army ants have of foraging in a self-organising way are used as metaphor to solve the task of searching, also in a self-organising way, for obstacles and trails in the robot’s visual field. The solution proposed in this thesis is to have multiple covert foci of attention operating as a swarm via pheromone-based interactions. This work represents the first embodied realisation of swarm cognition, which is a new born field that aims to uncover the basic principles of cognition by inspecting the self-organising properties of the collective intelligence exhibited by social insects. Hence, this thesis contributes to robotics as an engineering discipline, and to robotics as a modelling discipline supporting the study of adaptive behaviour.
|
|---|