Abstract: Robot swarms are often said to exhibit emergent properties. Yet, it is possible to design controllers with predictable outcome. We illustrate this for two canonical problems, multi-robot rendezvous and cooperative transport. The simplicity of the controllers (some do not even require arithmetic computation) facilitates their analysis. In the second part of the talk, we address the problem of inferring the rules of swarming agents through observation. We propose Turing Learning - the first system identification method not to rely on pre-defined metrics - and test it on a physical swarm of robots. Finally, we discuss novel development tools. We present OpenSwarm, an operating system for miniature robots, and formal methods for automatic code generation. We report on experiments with up to 600 physical robots.

Abstract: We present an overview of our recent work on self-organizing heterogeneous swarm systems that can show a wide variety of robust self-organizing spatio-temporal patterns. Our swarms consist of multiple types of very simple, kinetically interacting particles with no elaborate sensing, computation, or communication capabilities. We examine the effects of (1) heterogeneity of components, (2) differentiation/re-differentiation of components, and (3) local information sharing among components, on the self-organization of swarm systems, which are characterized using several kinetic and topological metrics. Results showed that (a) heterogeneity of components had a strong impact on the structure and behavior of the swarms, (b) dynamic differentiation/re-differentiation of components and local information sharing helped the swarms maintain spatially adjacent, coherent organization, (c) dynamic differentiation/re-differentiation contributed to the production of more diverse behaviors of swarms, and (d) stochastic re-differentiation of components also naturally realized a self-repair capability of self-organizing patterns. We also explore evolutionary methods to design novel, non-trivial self-organizing patterns, using either interactive evolutionary computation or spontaneous evolution within an artificial ecosystem. Finally, we demonstrate that these self-organizing swarm systems are remarkably robust against dimensional changes from 2D to 3D, although spontaneous evolution works more efficiently in a 2D space.

Abstract: Swarms of drones (but the discussion applies to any kind of swarm) are a promising paradigm because combining individuals offers much more features than increasing the capacity of a single entity (?The whole is more than the sum of its parts? Aristotle, Metaphysics). It also raises a number of issues among which failure. In most systems, failure is considered an exception. But when thousands of autonomous entities communicating with each other and adapting their behavior so as to achieve a global mission (this is referred to as swarm intelligence) are considered, the situation is quite different. Statistically, a number of individuals will fail and a number of messages will be lost (because of collisions, interferences, e tc.). Thus, ?In adaptive systems [...] classical separation between ?nominal operation" and "faults" becomes untenable; system is continuously operating under faults? [ Werner J.A. Dahm, Director, Security & Defense Systems Initiative, Arizona State University in his keynote at AIAA Guidance, Navigation, and Control Conference 19 - 22 August 2013, Boston , Massachusetts ]. Applications should then be built so that the failure of an individual entity does not imply the failure of the global mission. In other words, an entity should not rely on any expected behavior of the other entities of the swarm. As a consequence, a mission should be designed respecting the following principles: a global result (a global property to attain) should be targeted instead of a local individual result, which could not be guaranteed; it should be qualitative rather than quantitative, since the worst case is always possible; no individual can assume a peer in the swarm is present in its neighborhood; no individual can assume a peer in the swarm is lost (it can simply be temporarily unreachable); no communication can be assumed to get through. A mission is thus de facto designed as an emergent behavior/property at the global swarm level that results from local individual behaviors. Obeying the above principles makes it possible to achieve real world missions that can be ?guaranteed? resilient to individual failures and communications faults. The counterpart (but it is worth the cost) is that it often leads to bigger resource consumption.

Abstract: Social interaction increases significantly the performance of a wide range of cooperative systems, but natural swarms seem to limit the number of social connections. Flocking starlings interact on average with a fixed number of conspecifics and swarms of midges regulate their nearest-neighbor distance depending on the size of the swarm. This suggests that excessive social activity may have detrimental consequences. Using a canonical model of collective motion, we find that the responsiveness of a swarm is reduced when the social interaction exceeds a certain threshold. We find that the system can exhibit a large susceptibility even in the ordered phase (far from the critical point) if the amount of social interaction is set to an appropriate level. The same effect can be observed in collective decision-making models of distributed consensus, for example in a set of networked agents following the "majority vote" rule. If an external factor perturbs the state of a small sub-set of agents, this change will propagate through the network at a speed that depends on the number of social connections. These examples of distributed consensus show that an excess of social behavior can hinder their capacity to respond to fast perturbations. The result has far-reaching implications for the design of artificial swarms or interaction networks: even ignoring the costs of establishing connections and transmitting information between agents, it may be desirable to limit the number of connections in order to achieve a more effective dynamical response.

Abstract: Fast transfer of information in groups can have survival value. An example is the so-called wave of agitationobserved in groups of animals of several taxa under attack. It has been shown to reduce predator success. It usually involves the repetition of a manoeuvre throughout the group, transmitting the information of the attack quickly, faster than the group moves itself. The specific manoeuvre underlying a wave is typically known, but not so in starlings (Sturnus vulgaris). Although waves of agitation in starling flocks have been suggested to reflect density waves, exact escape manoeuvres cannot be distinguished because flocks are spatially too far away. Therefore, waves may also reflect orientation waves (due to escape by rolling). In the present study, we investigate this issue in a computational model based on self-organization, StarDisplay, because its flocks resemble empirical data in many respects. The model comprises moving individuals that fly based on fixed-wings aerodynamics, coordinate with 6-7 of their closest neighbours and try to stay close to the site for sleeping. We use this model because its flocks have been shown to resemble starling flocks in many traits, such as variability of shape, diffusion in the flock and behavior during turning. In the model, we show that agitation waves result from changes in orientation rather than in density. They resemble empirical data both qualitatively in visual appearance and quantitatively in wave speed. In the model, local interactions with only two to seven closest neighbours suffice to generate empirical wave speed. Wave speed increases with the number of neighbours repeated from and the distance to them. It decreases with reaction time and with time to identify the escape manoeuvre of others and is not affected by flock size. Our findings can be used as predictions for empirical studies.

Abstract: Collective animal groups exhibit spontaneous ordered movement patterns with robustness against environmental perturbations (e.g., predator attacks), which would be realized by perpetual propagations of information through interactions. Recently it has been reported that the internal structures of collective animal groups are not fixed in time. Such diverse movements of individuals, which might be expected to be detrimental for collectivity, play a prominent role in facilitating interactions with various neighbors and thereby robust collective motion and information transfer. In the present study, we first address whether there is a balance between excessive movement that is detrimental to the maintenance of the group and movement that is too slight to contribute to collectivity. We previously found that in fish schools even though individuals show schooling behavior with high polarity, each individual movement relative to the center of the mass of the group displays Le?vy walk pattern in which the step-length follows power-law distribution. Here, we carry out a simple simulation of individuals? movement within the school, derived from above experimental data. In this simulation, we set an area whose diameter is the mean school size as an internal area of the school, and distribute agents in the area so as to have same density with the school. Movement pattern of each agent obeys one of the following; Le?vy walk, Brownian walk, and Ballistic movement. When we compare efficiencies to contact with neighbors and of information transfer between movement patterns, we find that Le?vy walk that was observed in real school is optimal to interact with various neighbors and thereby robust information transfer. Second, we propose a new computational model of collective behavior in which diverse individual moves positively contribute to form a group, and show that this model can realize densely collective motion with high polarity, showing inherent turbulent motion appearing as a Le?vy walk pattern.

Abstract: Atta leaf-cutter ant colonies reach sizes of many million individuals. Achieving such colony sizes is owed in part to the capacity of foraging on leaves, a virtually unlimited resource in their natural habitat. Nutrients from this low?quality resource are made available to the colony via a symbiotic fungus.While foraging on leaves always follows approximately 10 hour periods, the timing of their foraging bouts is notoriously perplexing: colonies shift seemingly randomly between diurnal and nocturnal foraging, and even within a colony different patterns may co?occur on different foraging trails. What governs these rhythms of foraging activity within colonies? The mechanisms are poorly known, and we have made observations further adding to the mystery: the periodicity of foraging cycles can vary for harvesting different substrates, and multiple cycles can co?exist on the same foraging trail.We propose the existence of a control mechanism that regulates foraging activity to satisfy a target intake rate. Processing of leaf fragments and integration into the fungus is time-consuming ? rather than matching intake rate to processing rate, leaf-cutter ants allow leaf fragment buildup inside the nest. These leaf caches can function as warehouse levels, and have the potential to serve as an inhibitory regulatory signal in a foraging control mechanism.Some forage material however, like fallen fruit, is directly consumed by the adults and requires less handling time. In our field experiments, we find foraging on fruits to be continuous, implying the absence of inhibitory feedback for this resource type.

Abstract: It has been proposed that artificial swarm design can be tackled from the angle of interaction network design. Network science provides a very powerful framework allowing bridging the gap between local dynamics and interactions at the agents level and global response at the swarm level. In the context of swarm dynamics, identifying emerging patterns and their associate properties?especially for swarms lacking apparent order?are known to be very challenging. Through specific network-theoretic analyses, ?hidden? structures emerging through self-organization can be uncovered. For instance, increasing the network degree usually improves the performance of swarms, but it is known that most natural swarms operate with a limited number of social connections. Here, we focus on studying the influence that the interaction network topology has on the dynamical response of a swarming system. We present the consequences of such excessive social behaviors on some dynamical properties of a wide range of networked systems with different topologies.