Robustness of interaction control in robot swarms

Abstract

Robotics is making significant strides in its capabilities and is being integrated into many real-world tasks. Some of these are carried out in unpredictable and dynamic environments, for example, search and rescue operations in a disaster area. Other problems, such as medical applications that require the robots to be very small, have led to significant progress in nanorobotics, in which the processor and actuating abilities are limited due to physical constraints. To solve these problems, we must look to methods that are robust to changes in the environment as well as solutions that draw maximum effect from the abilities of the robot. This may be achieved by utilising the potential of a group of robots often related to the principle of swarm intelligence. In this context, we study emergent phenomena in a group of agents each having simple low-level rules and strongly rely on local communication, resulting in increased flexibility and versatility. Individually the agents cannot perform well. As a collective, however, they are able to complete complex tasks and display implicit cognitive abilities that are critical in the applications of swarm robotics. This Thesis aims to put forward the argument for the strengths of swarm intelligence, and furthermore how it may be incorporated in applications. This Thesis first looks at a swarm having the aim to maximise energy consumption in a fixed-sized environment. The agents can sense the whole environment and augment sensory information with local communication as input for a neural network which is trained by an evolutionary strategy. Although successful for one source, as the number of sources increases frustration builds within the swarm and its effectiveness declines. It is shown that by limiting information through evolving the physical properties, the swarm can avoid this confusion. We have studied models and techniques that do not require explicit learning, and that can adapt as the situation changes. The Thesis first looks at reaction-diffusion equations being deployed onto a swarm and form Turing patterns, which are stable periodic patterns usually either honey-comb spotted patterns or stripes. We then explore how patterns can be induced via the environment. This is deployed onto a virtual Kilobot swarm to demonstrate how Turing patterns can guide robotic systems, even within the limitations of the Kilobot platform. We then present another system which has the aim to separate the agents in the swarm to allow maximum coverage of the environment. Starting with a one-dimensional dynamical system we show how this system operates given a periodic boundary condition. Using this as a foundation, we include a fixed environment, limited range of communication, and finally how the rules can be employed onto individual agents in a two-dimensional environment for maximal coverage. We compare this to the separation rule in classical swarm systems and show that our system has a smoother distribution allowing it to cover the environment quicker. This can then be taken further to circumvent obstacles or surround objects using the principle of separation.