In this thesis, we use evolutionary robotics techniques to automatically design and synthesise<p>behaviour for groups of simulated and real robots. Our contribution will be on<p>the design of non-trivial individual and collective behaviour; decisions about solitary or<p>social behaviour will be temporal and they will be interdependent with communicative<p>acts. In particular, we study time-based decision-making in a social context: how the<p>experiences of robots unfold in time and how these experiences influence their interaction<p>with the rest of the group. We propose three experiments based on non-trivial real-world<p>cooperative scenarios. First, we study social cooperative categorisation; signalling and<p>communication evolve in a task where the cooperation among robots is not a priori required.<p>The communication and categorisation skills of the robots are co-evolved from<p>scratch, and the emerging time-dependent individual and social behaviour are successfully<p>tested on real robots. Second, we show on real hardware evidence of the success of evolved<p>neuro-controllers when controlling two autonomous robots that have to grip each other<p>(autonomously self-assemble). Our experiment constitutes the first fully evolved approach<p>on such a task that requires sophisticated and fine sensory-motor coordination, and it<p>highlights the minimal conditions to achieve assembly in autonomous robots by reducing<p>the assumptions a priori made by the experimenter to a functional minimum. Third, we<p>present the first work in the literature to deal with the design of homogeneous control<p>mechanisms for morphologically heterogeneous robots, that is, robots that do not share<p>the same hardware characteristics. We show how artificial evolution designs individual<p>behaviours and communication protocols that allow the cooperation between robots of<p>different types, by using dynamical neural networks that specialise on-line, depending on<p>the nature of the morphology of each robot. The experiments briefly described above<p>contribute to the advancement of the state of the art in evolving neuro-controllers for<p>collective robotics both from an application-oriented, engineering point of view, as well as<p>from a more theoretical point of view. / Doctorat en Sciences de l'ingénieur / info:eu-repo/semantics/nonPublished

We look at robotic systems made of separate discrete components that, by self-assembling, can organize into physical structures of growing size. We review 22 such systems, exhibiting components ranging from passive mechanical parts to mobile<p>robots. We present a taxonomy of the systems, and discuss their design and function. We then focus on a particular system, the swarm-bot. In swarm-bot, the components that assemble are self-propelled modules that are fully autonomous in power, perception, computation, and action. We examine the additional capabilities and functions self-assembly can offer an autonomous group of modules for the accomplishment of a concrete task: the transport of an object. The design of controllers is accomplished in simulation using<p>techniques from biologically-inspired computing. We show that self-assembly can offer adaptive value to groups that compete in an artificial evolution based on their fitness in task performance. Moreover, we investigate mechanisms that facilitate the design of self-assembling systems. The controllers are transferred to the physical swarm-bot system, and the capabilities of self-assembly and object transport are extensively evaluated in a range of different environments. Additionally, the controller for self-assembly is transferred and evaluated on a different robotic system, a super-mechano colony. Given the breadth and quality of the results obtained, we can say that the swarm-bot qualifies as the current state of the art in self-assembling robots. Our work supplies some initial evidence (in form of simulations and experiments with the swarm-bot) that self-assembly can offer robotic systems additional capabilities and functions useful for the accomplishment of concrete tasks.<p> / Doctorat en Sciences de l'ingénieur / info:eu-repo/semantics/nonPublished

In the context of evolutionary robotics and neurorobotics, artificial neural networks, used as controllers for animats, are examined to identify principles of neuro-control, network organization, the interaction between body and control, and other likewise properties. Before such an examination can take place, suitable neuro-controllers have to be identified. A promising and widely used technique to search for such networks are evolutionary algorithms specifically adapted for neural networks. These allow the search for neuro-controllers with various network topologies directly on physically grounded (simulated) animats. This neuro-evolution approach works well for small neuro-controllers and has lead to interesting results. However, due to the exponentially increasing search space with respect to the number of involved neurons, this approach does not scale well with larger networks. This scaling problem makes it difficult to find non-trivial, larger networks, that show interesting properties. In the context of this thesis, networks of this class are called mid-scale networks, having between 50 and 500 neurons. Searching for networks of this class involves very large search spaces, including all possible synaptic connections between the neurons, the bias terms of the neurons and (optionally) parameters of the neuron model, such as the transfer function, activation function or parameters of learning rules. In this domain, most evolutionary algorithms are not able to find suitable, non-trivial neuro-controllers in feasible time. To cope with this problem and to shift the frontier for evolvable network topologies a bit further, a novel evolutionary method has been developed in this thesis: the Interactively Constrained Neuro-Evolution method (ICONE). A way to approach the problem of increasing search spaces is the introduction of measures that reduce and restrict the search space back to a feasible domain. With ICONE, this restriction is realized with a unified, extensible and highly adaptable concept: Instead of evolving networks freely, networks are evolved within specifically designed constraint masks, that define mandatory properties of the evolving networks. These constraint masks are defined primarily using so called functional constraints, that actively modify a neural network to enforce the adherence of all required limitations and assumptions. Consequently, independently of the mutations taking place during evolution, the constraint masks repair and readjust the networks so that constraint violations are not able to evolve. Such functional constraints can be very specific and can enforce various network properties, such as symmetries, structure reuse, connectivity patterns, connectivity density heuristics, synaptic pathways, local processing assemblies, and much more. Constraint masks therefore describe a narrow, user defined subset of the parameter space -- based on domain knowledge and user experience -- that focuses the search on a smaller search space leading to a higher success rate for the evolution. Due to the involved domain knowledge, such evolutions are strongly biased towards specific classes of networks, because only networks within the defined search space can evolve. This, surely, can also be actively used to lead the evolution towards specific solution approaches, allowing the experimenter not only to search for any upcoming solution, but also to confirm assumptions about possible solutions. This makes it easier to investigate specific neuro-control principles, because the experimenter can systematically search for networks implementing the desired principles, simply by using suitable constraints to enforce them. Constraint masks in ICONE are built up by functional constraints working on so called neuro-modules. These modules are used to structure the networks, to define the scope for constraints and to simplify the reuse of (evolved) neural structures. The concept of functional, constrained neuro-modules allows a simple and flexible way to construct constraint masks and to inherit constraints when neuro-modules are reused or shared. A final cornerstone of the ICONE method is the interactive control of the evolution process, that allows the adaptation of the evolution parameters and the constraint masks to guide evolution towards promising domains and to counteract undesired developments. Due to the constraint masks, this interactive guidance is more effective than the adaptation of the evolution parameters alone, so that the identification of promising search space regions becomes easier. This thesis describes the ICONE method in detail and shows several applications of the method and the involved features. The examples demonstrate that the method can be used effectively for problems in the domain of mid-scale networks. Hereby, as effects of the constraint masks and the herewith reduced complexity of the networks, the results are -- despite their size -- often easy to comprehend, well analyzable and easy to reuse. Another benefit of constraint masks is the ability to deliberately search for very specific network configurations, which allows the effective and systematic exploration of distinct variations for an evolution experiment, simply by changing the constraint masks over the course of multiple evolution runs. The ICONE method therefore is a promising novel evolution method to tackle the problem of evolving mid-scale networks, pushing the frontier of evolvable networks a bit further. This allows for novel evolution experiments in the domain of neurorobotics and evolutionary robotics and may possibly lead to new insights into neuro-dynamical principles of animat control.