The autonomous navigation of a mobile robot in an unknown environment is a complex task that raises many problems related to perception, categorization, planning and motor control. Solving all these problems in an integrated way remains a challenge for roboticians. The approach developed in the laboratory since its creation is original in that it relies largely on the most recent discoveries in neurology. We will demonstrate in this thesis that models from hypocampal and parahympocampal system studies allow the robot to navigate and situate itself in unknown environments. For this purpose, we will rely on an internal laboratory tool: Promethe. The latter is a development environment based on middleware that manages both low-level robotic aspects such as speed and higher-level aspects such as neural network learning.

Applied deep learning to elements of the highway landscape (traffic signs, pedestrians and vehicles) is a crucial need for autonomous vehicle. During this intership, I use the Caffe library (and its Python overlay) to perform image learning and recognition. The early stages of the intership consisted in a discovery of computer vision theory, neural networks and the library. I then conducted several experiments in order to better understand how a convolutional network works and ultimately develop scripts to detect elements in an image. The application phase then began with the learning of several models, more and more precise (by discovering the effects of learning parameters on precision). Thus, a reliable panel model was created (95% accuracy on classification) as well as a car model (94% accuracy).

The design and characterization of chemical microsensor dedicated to the detection of gaseous species is an important part of air quality evaluation. Indeed, closed environments accumulate quantities of puluants and detect them, even more when there are higly toxic at very low dose, is a major public health issue.