Abstract : The goal of this research is to develop a behavior oriented design methodology dedicated to mobile and autonomous robotics. The sought design methodology should start from what is known, i.e. a description of the mission to be fulfilled. Noticing that animals have a high level of autonomy that would provide to robots the ability to fulfill numerous missions, we have focused our search, in the frame of the animat approach, on the main mechanism that has built and designed living creatures: natural selection. One of the features of natural selection, is that it only takes into account the result, i.e. the ability to transmit the genes. In an artificial context, such an algorithm will take into account the ability to solve the given task. We have focused our research on the scaling problem in terms of the complexity of behaviors exhibited by evolved neuro-controllers – both structure and parameters of a neural network being generated by the evolutionary algorithm. A study on modularity has revealed us the importance of the selection pressure definition, leading to a so called exaptation approach. We propose then the use of a multi-objective scheme in which the fitness function rewarding task achievement is associated to objectives that may be independent from this goal. In this “multi-objectivization” context and besides work on exaptation, our contributions are the following: * multi-objective incremental approach: each sub-task is associated to a separate objective. It is then neither necessary to weight them, nor to choose a resolution order, nor to decide when to switch from one task to another; * behavior diversity: an objective measuring the mean distance in the space of behaviors to the rest of the population is used to maintain a high diversity in this space. This approach has led to very efficient results, even with the most simple neural network encodings; * transferability: a transferability objective is maximized to facilitate the transfer from simulation to reality without any significant performance loss. In the perspective of designing more cognitive controllers and in the frame of the ANR EvoNeuro project, we have proposed a methodological approach relying on neurosciences and consisting in generating neural networks close to models developed in this research field. On the basis of a specific neural network encoding, we have generated neural networks exhibiting simple action selection or working memory abilities. The goal of this project is to generate neural network showing other cognitive abilities studied in computational neuroscience while relying on their evaluation protocol and then applying these results in a robotic context. In this project, we have also developed a multi-objective analysis method allowing neuroscientists to compare and analyze their own models on the basis of a multi-objective optimization.