The simplicity of the individual robots in a collective is another of the features emphasised in a swarm intelligence-based approach, as it is in behaviour-based robotics. The use of simple robots can result in a less costly system, and one that is more robust in the sense that simple robots are less likely to fail (since there is less to go wrong). Another advantage of simplicity is the ability to respond rapidly and flexibly to changes in the environment. The kind of simplicity we have in mind refers in particular to the control system used. Individual robots in a collective can themselves be subject to reactive, or deliberative control. A reactive architecture is one that “tightly couples perception to action without the use of intervening abstract representations” (Arkin, 1998). A deliberative architecture, on the other hand, relies on abstract representations of the world. Between the two lie those systems that extend purely reactive systems with
some memory capabilities; in other words, rather than just reacting to a stimulus, the robots are affected by an internal register which has some form of memory.
The earlier formulations of behaviour-based robotics stressed the importance of reactivity (Brooks, 1986, 1991), and its advantages in terms of the ability to respond quickly to stimuli in the environment. More recent formulations of behaviour-based robotics have incorporated some degree of memory and representation needed to accomplish more complex tasks, while maintaining an emphasis on avoiding the use of centralised representations and control (Mataric, 1997b). Arkin (1990) advocated the use of control systems consisting of a hybrid of reactive and deliberative control. However, the situations in which deliberative control is likely to be the preferred option are those in which uncertainty is limited, and the world can be accurately modelled, not the kinds of situation for which swarm robotics are best suited. An emphasis on the application of swarm intelligence principles to collective robotics implies the use of control systems that are as simple and reactive as possible.
The simplest control system for an individual robot, then, is one in which control is as close as possible to sensors and actuators as, for example, when an artificial neural network is used and the inputs are stimulated by raw sensor values, while the outputs control motor speed and direction. At a higher level, the architecture can be organised into basic behaviours, each representing a perception-action loop. The starting point is usually behavioural modules responsible for robot movement (for example, a wander module, and an obstacle avoidance module). Higher level modules responsible for finding objects or moving towards a goal can then be added, depending on the task in question. Decisions about the design of behavioural modules are usually the responsi- bility of a human designer. Once a set of behavioural modules is chosen, some method of combining them is required. Behavioural modules can be combined by means of either a selection method (switching control to the most appropriate module), or by fusing them. We shall consider examples of each of these in turn.
Probably the best known selection method is that represented by Brooks’ subsumption architecture. In a subsumption architecture (Brooks, 1986), a fixed priority scheme is defined for basic behaviours such that enabling one of them results in the suppression or inhibition of others, so that only one behaviour is active at any one time. An alternative switching mechanism was proposed by Maes (1989) based on spreading activation between modules. A more recent development of Maes’ system was proposed by Jung and Zelinsky (1999): a selection method termed architecture for behaviour-based agents, (ABBA). The selection mechanism is based on a winner-take-all scheme. Activation is spread among competence modules on the basis of the output of feature detectors, and the pre-conditioning competence module. When the activation level of a competence module reaches threshold, it becomes active. The scheme has been tested on two heterogeneous Yamabico robots (http://www.roboken.esys.tsukuba.ac.jp/ english/Yamabico) performing a collaborative cleaning task.
Switching and selection methods rely on the assumption that only one behavioural module should be active at any one time. The alternative approach is to adopt some form of fusion of modules, where the outputs of several active modules are fused, or combined in some way, to result in a single behaviour that reflects the influence of several modules. For example, under the motor schema-based approach (Arkin, 1989), primitive behaviours, or motor schemes, can be active simultaneously, and combined cooperatively. Behaviour
is obtained by multiplying the vector response of each motor schema by a gain, and then summing and normalizing the result. The DAMN architecture (Payton, Rosenblatt, & Keirsey, 1990) used by researchers at Carnegie-Mellon University for controlling unmanned ground vehicles, similarly relies on fusion, using a scheme by which each behaviour votes for and against each of a set of possible vehicle actions, and an arbiter performs command fusion to select the most appropriate action.
To summarise: The application of swarm intelligence to collective robotics implies the need for simple robots that can respond rapidly and flexibly to the environment. The main way to achieve this, at present, is to rely on a system of reactive control at the individual level, or a set of reactive behavioural modules combined through some form of action selection.