Robots are usually designed following an internal organization of the kind "sense-plan-act" proposed during the early years of research in Artificial Intelligence. It involves the preliminary development of interaction models between the robot and its environment and of sensory data processing algorithms leading eventually to the generation of motor commands. In this specific context, the features of the environment perceived by the robot fundamentally result from the sensory data processing implemented by the robotician. They can thus turn out to be inadequate to capture the real sensorimotor interaction the robot has with the environment. Another approach, developed in this thesis, consists in redesigning the perception problematic in robotics. It is inspired by the sensorimotor contingencies theory that proposes considering our perception not as a phenomenon happening in our brain but as an interaction we have with the environment. This point of view, radically opposed to classical postulates, induces that perceiving is not innate but is learned through the discovery of sensorimotor relations that underlie our experience of the world. The goal of this thesis is to apply this new paradigm in the field of robotics. More precisely, this work aims at determining how a naive robot can discover and characterize the space in which his body and the environment are immersed through the only analysis of its sensorimotor data. To answer this question, an approach is developed on the basis of the compensability of sensory variations generated by the displacements of the robot/environment system, a concept initially introduced by H.Poincaré. It allows our robots to determine the dimension of the external geometrical space and to build an internal representation of it allowing in the long term to interpret their experience and to guide their action.
