On behalf of:

Abstract: The swarm intelligence literature frequently asserts that swarms exhibit high levels of robustness. That claim is, however, rather less frequently supported by empirical or theoretical analysis. But what do we mean by a ‘robust ’ swarm? How would we measure the robustness or – to put it another way – fault-tolerance of a robotic swarm? These questions are not just of academic interest. If swarm robotics is to make the transition from the laboratory to real-world engineering implementation, we would need to be able to address these questions in a way that would satisfy the needs of the world of safety certification. This paper explores fault-tolerance in robot swarms through Failure Mode and Effect Analysis (FMEA) and reliability modelling. The work of this paper is illustrated by a case study of a wireless connected robot swarm, employing both simulation and real-robot laboratory experiments.

Initiated by N. Reeves, the Mascarillons project stands at a crossroad between Art and Science1. It aims to bring together researchers in both artistic and scientific domains to collaborate towards the production of a robotic environment dedicated to architectural research, with a major potential for multi-media performance. Because of the tight constraints of the project, a multi-level design methodology is proposed, consisting of the parallel development of real robots and increasingly abstract modeling tools. The interaction of these experimental levels is expected to hasten the progress towards an efficient solution by taking every technological and physical constraint into account. 1. THE ARTISTIC PROJECT Taking its roots in numerous biological examples, the swarm

Abstract — In swarm robotic systems emergent swarm properties are particularly difficult to analyse and model. This paper describes a simple but effective algorithm for emergent swarm taxis (swarm motion toward a beacon) in a 2D or 3D wireless connected swarm of minimalist mobile robots. The paper then undertakes a deep analysis of the swarm taxis by identifying both first and second order micro-level robot interactions and quantifying the contribution of each such interaction to the macro-level swarm behaviour. From the analysis we develop a simple quantitative model that is able to predict swarm velocity with reasonable accuracy. Although the analysis is specific to the swarm algorithm in question, we believe that the methodology presented has generic value to swarm modellers. I. BACKGROUND. The fields of robotics and artificial intelligence have seen

Designing and analyzing self-organizing systems such as robotic swarms is a challenging task even though we have complete knowledge about the robot’s interior. It is difficult to determine the individual robot’s behavior based on the swarm behavior and vice versa due to the high number of agent–agent interactions. A step towards a solution of this problem is the development of appropriate models, which accurately predict the swarm behavior based on a specified control algorithm. Such models would reduce the necessary number of time-consuming simulations and experiments during the design process of an algorithm. In this paper we propose a model with focus on an explicit representation of space because the effectiveness of many swarm robotic scenarios depends on spatial inhomogeneity. We use methods of statistical physics to address spatiality. Starting from a description of a single robot we derive an abstract model of swarm motion. The model is then extended to a generic model framework of communicating robots. In two examples we validate models against simulation results. Our experience shows that qualitative correctness is easily achieved while quantitative correctness is disproportionately more difficult but still possible. 1

By compiling macroscopic models we analyze the adaptive behavior in a swarm of autonomous robots generated by a bio-inspired, distributed control algorithm. We developed two macroscopic models by taking two different perspectives: A Stock & Flow model, which is simple to implement and fast to simulate, and a spatially resolved model based on diffusion processes. These two models were compared concerning their prediction quality and their analytical power: One model allowed easy identification of the major feedback loops governing the swarm behavior. The other model allowed analysis of the expected shapes and positions of observable robot clusters. We found a high correlation in the challenges posed by both modeling techniques and we highlighted the inherent problems of inferring emergent macroscopic rules from a microscopic description of swarm behavior. Key words: macroscopic modeling, swarm robotics, bio-inspired robotics 1

First and for most, I would like to thank my supervisor Prof. Alan FT Winfield for his guide and advise to complete this work. I really appreciate the freedom that Alan gave me in choosing the research direction and method. Along the way I have benefited a lot from the discussion with him, both from formal supervision meeting and informal research chatting every Friday lunch time. I would like especially to thank Alan for the help to correct all the grammars in English through the whole thesis with great patient. Without the help from Alan, this thesis couldn’t reach its final form. I am also grateful to my second supervisor Dr. Jin Sa for the insightful discussion about the thesis and the project. I want to thank Jin for personally supporting me in settling down in Bristol at the beginning of my study, which makes the life much easier. I would like to thank the director of the Bristol Robotics Laboratory, Prof. Chris Melhuish for providing an extremely friendly and stimulating research environment. I would like also to thank all the colleagues in the lab for all the suggestions and kindless help during last three years. A special thank goes to Jan Dyre Bjerknes for the useful and helpful discussion in swarm robotics, and for his organisation of all kinds of parties and activities.

distributed safety-critical systems. Swarm Intelligence provides us with a powerful new paradigm for building fully distributed de-centralised systems in which overall system functionality emerges from the interaction of individual agents with each other and with their environment. Such systems are intrinsically highly parallel and can exhibit high levels of robustness and scalability; qualities desirable in high-integrity distributed systems. Making use of a laboratory based swarm robotic system as a case study, this review paper explores dependability, robustness and reliability modelling in swarm based systems, and argues that there is considerable merit in further investigating their application to distributed safety-critical systems. 1

It is a characteristic of swarm robotics that specifying overall emergent swarm behaviours in terms of the low-level behaviours of individual robots is very difficult. Yet if swarm robotics is to make the transition from the laboratory to real-world engineering realisation we need such specifications. This paper explores the possibility of using temporal logic to formally specify, and possibly also prove, the emergent behaviours of a robotic swarm. The paper makes use of a simplified wireless connected swarm as a case study with which to illustrate the approach. Such a formal approach could be an important step toward a disciplined design methodology for swarm robotics. 1.

The search for generally applicable methods in swarm intelligence aims to gain new insights about natural swarms and to develop design methodologies for artificial swarms. The ideal would be a ‘swarm calculus ’ that allows to calculate key features such as expected swarm performance and robustness on the basis of a few parameters. A path towards this ideal is to find methods and models that have maximal generality. We report two models thatmight be examples of exceptional generality. First, we present an abstract model that describes the performance of a swarm depending on the swarm density based on the dichotomy between cooperation and interference. Second, we give an abstract model for decision making that is inspired by urn models. A parameter, that controls the feedback based on the current consensus, allows to understand the effects of an increasing probability for positive feedback over time in a decision making system. 1