Abstract : We conducted research mainly in the areas of navigation, perception and learning for mobile robots. These studies, oriented toward a cognitive approach to robotics have the overall goal of allowing robots to adapt to their environment, providing ba- sic primitives such as open space, position, or the presence of objects necessary to choose actions. A large part of this work is inspired by capabilities found in nature, but without trying to reproduce exactly the biological systems inner functioning. Navigation, especially mapping, has so far been the focus of our work. During our PhD thesis, we developed a mapping method based on a Bayesian filter, applied to map structure and perceptions inspired from biological knowledge about the naviga- tion capabilities of the rat (Filliat and Meyer, 2000; Filliat, 2001; Filliat and Meyer, 2002). The advantage of this approach is to allow, from simple sensors, global lo- calization during mapping, providing a good robustness to kidnapping, at the price of a relatively slow exploration. This biological inspiration then disappeared during our work at the Direction Générale pour l'Armement where we participated in the development of a demonstrator using conventional mapping techniques based on la- ser ranging and obstacle avoidance in dynamic environments (Dalgalarrondo et al., 2004). Since 2005, at ENSTA ParisTech, our work has focused on the problem of topological navigation. We developed a learning-based topological approach (Filliat, 2007; Filliat, 2008) in which a user shows the rooms to be recognized and shows the path between the different rooms. We have also developed a topological mapping ap- proach using a loop closure detection algorithm that can detect the return of a robot to previously visited locations (Angeli et al., 2008a; Angeli et al., 2009; Bazeille and Filliat, 2011). Finally, this work has now expanded since 2009 as part of the PACOM (Filliat et al., 2009) project to the problem of semantic mapping (Jebari et al., 2011). The objective is to obtain models of the environment that include higher-level infor- mation, in particular information closer to those used by humans, such as rooms or objects found in the environment. In terms of perception, some of these works have used laser ranging, which is well suited to navigation. However, they are focused primarily on the use of vision. In particular, we investigated the problem of representation of visual information that is essential to provide robustness to noise while providing necessary information to ap- plications. We have developed an incremental approach inspired by the "bag of visual words" model, which we applied to qualitative localization (Filliat, 2007), topological localization (Angeli et al., 2008b), visual guidance (Filliat, 2008) and object recog- nition (Rouanet et al., 2009). In collaboration with Pierre-Yves Oudeyer we have extended this representation to the auditory (Mangin et al., 2010) and audio-visual (ten Bosch et al., 2011) recognition. We also studied the problem of active perception in order to improve localization capabilities (Filliat and Meyer, 2000; Filliat, 2007) and improve the robustness of visual guidance (Filliat, 2008). Finally, most of these studies have used learning methods to provide adaptability to localization, mapping or object recognition. We mainly worked with Bayesian mo- dels and we have developed active methods, allowing a robot to select the training samples to improve its performance. These methods can also benefit from interac- iv tions with the user to adapt the concepts learned by the robot. We applied them to the recognition of rooms (Filliat, 2007; Filliat, 2008) and objects (Rouanet et al., 2009). We also applied non-negative matrix factorization, an unsupervised learning method, for audio-visual recognition of objects (ten Bosch et al., 2011). This last application was developed in the context of developmental robotics, where we take inspiration from humans to create intuitive and long term learning methods, an approach we are currently developing within the MACSi project (Sigaud et al., 2010). In line with this work, we plan to continue our research on the topic of semantic navigation and learning for perception in the context of developmental robotics. This research will seek to provide rich environmental models to the robot or its user. These models will contain useful information for analysing the situation and for the robot tasks. These methods will be mainly applied in the context of navigation in indoor or urban environments and to service or assistive robotics, in direct interaction with humans.