|Keywords Swarm Robotics, Self-Organised Robot Systems
In software engineering, a design pattern associates a particular class of known problem with a particular class of effective solution. Swarm robot engineers would benefit from similar design patterns that each associate specific robot control schemes with desired collective performance. Here, we characterise such design patterns for robot swarms, initially in the context of collective foraging.
In the setting that we explore there are three main factors that affect foraging performance: recruitment, interference and swarm size. Recruitment (i.e., the utilisation of shared social information) promotes exploitation of known deposits, and is effective when deposits are hard to find but contain significant resource. When these conditions do not hold, interference between robots can prevent large groups from foraging effectively. Since more communication takes place in larger swarms, the effect of recruitment is more pronounced. Alternative control schemes perform equivalently in small swarms but diverge in larger swarms.
We characterise these trends in terms of information flow within swarms (via scouting and recruitment) and the costs associated with converting information into foraging success. We introduce Information Value as a measure of the relative change in foraging success that results from a new piece of information.
Because they exist in a physical, spatially extended, dynamic world, robots cannot turn information into reward for free: a robot travelling to a resource site pays a “misplacement cost”; a robot foraging from an exhausted or a crowded site pays an “opportunity cost” (especially important in dynamic environments where deposit locations and rewards change over time); a robot waiting for teammates to arrive in order to complete a task pays a “coordination cost”.
Here we exploit Information Value analysis to gain understanding of the way in which robots explore the environment and share information and how this affects their suitability for a given spatio-temporal distribution of resources in order to fuel the creation of design patterns and associate them with appropriate collective tasks.
For example, an Information Exchange Centre pattern, according to which robots may only exchange information in a designated area, promotes spatio-temporal coordination but can inhibit exploration, whereas an Opportunism pattern that ensures robots always exploit a higher value deposit when they learn about it from their teammates, leads to rapid preferential completion of high-reward tasks, but can cause poor sampling of the arena.
By moving beyond the reporting of successful swarm performance on a specific task to a design-pattern-oriented analysis of alternative control schemes, we are able to take steps towards more general-purpose design principles for robot swarm applications.