Deep Learning is a machine learning technique that seeks to define neural networks based on pattern recognition from input data, and to achieve results with a high level of confidence and time efficiency. In this paper, this technique is used in two processes that aim to improve the indoor exploration task of a mobile robot. The first process enables the recognition of physical structures such as stairs, elevators and corridors, based on images obtained from an on-board camera. The second process identifies the proximity to obstacles (based on the information from an on-board laser sensor) in order to determine the direction and moving speed of the robot for navigating an indoor space. This work stands out because it uses Deep Learning in both processes, through the implementation of a parametrizable generic neural network system, to create multiple network instances that are customizable according to the type of information, and to integrate them with a mobile robot that navigates in simulated and experimental environments.