this thesis describes part of the work carried out in the European project Symbrion 1. This project aims at carrying out complex tasks requiring the cooperation of multiple robots in essaim robotics (at least 100 robots operating together). Many problems are being studied by the project, including: the self-assembly of robots into complex structures and the self-organisation of a large number of robots in order to carry out a common task. The main topic is the self-adaptation mechanisms for modular and essaim robotics, with an interest in strong coordination and cooperation capabilities at the spin-scale. The difficulties encountered in carrying out this project are due to the use of robots in open environments that remain unknown until the deployment phase. As the operating conditions cannot be predicted in advance, e-learning algorithms should be used to develop the behaviours used. When a large number of robots are used, several considerations need to be taken into account: reduced communication capacity, low memory, low computing capacity. Therefore, e-learning algorithms need to be distributed through the essay. Multiple approaches have already been proposed to address the challenges posed by decentralised online learning of robotic behaviours, including probabilistic robotics, reinforcement learning, and evolutionary robotics. However, the problem addressed in this thesis is characterised by the consideration of a group of robots (instead of a single robot). Moreover, due to the open nature of the environment, it is not possible to assume that the human engineer has the necessary knowledge to define the necessary elements for learning processes. Ensuring the integrity of the essaim is placed as the first element of a roadmap to define a set of steps necessary for a group of robot to perform a task in an open environment: — Step 1: Ensure the integrity of the saim. — Step 2: Keep robots available as a service to the user. — Step 3: Carrying out the task defined by the user. In the context of this thesis, we are working on step 1 of this roadmap and assume the following working hypothesis: In a context of collective robotics in an open environment, the performance of a task defined by the user involves first of all self-adaptive behaviour. The subject of this thesis is the realisation of decentralised algorithmic solutions that can guarantee the tegrity of a robot essaim in an open environment when a collective robotics system uses local communication. The main difficulty in its resolution is the need to take into account the ring-fencing. Depending on the common environment, robots may have to demonstrate a wide variety of global behaviours such as cooperation, specialisation, altruism or division of labour. In this thesis, we introduce and define the problem of environmentally-led Distributed Evolutionary Adaptation. We propose an algorithm to solve this problem. This algorithm has been validated both in simulation and on real robots. It has been used to study the problem of self-adaptation in the following environments: — Environment where the emergence of behavioural consensus is needed. — Approaches where robustness to environmental change is needed. — Areas where altruistic behaviour is needed.