• artificial intelligence
• autonomous vehicle
• convolutional neural network
• deep neural network
• global localization
• robotics

From the dawn of this two fields, Artificial Intelligence and Robotics have shared a bound that is very effective, you can not talk about one without mentioning the other. The AI models are flexible and robust so they can be integrated in a robot architecture to make it more performing and reliable. One of the fundamental pillars of Robotics is navigation: robot needs to navigate inside an environment in order perform its application tasks. However, to be able to do that it needs to localize itself inside of it. When outdoor, GPS is a reliable solution to provide absolute positioning. On the other hand, in indoor scenarios, the robot needs to create a map of the environment and use it to localize itself during the mission. In particular the initialization process of the localization is delicate, to the point that could mean success or fail of the mission. Nowadays current localization approaches, requires an initial estimate of the position which is difficult to have in fully autonomous scenarios which do not have human interaction with the robot. Several approaches already exist, however they require high computation and sometimes produce estimates that has a marked error. This study presents a Deep Neural Network model that is trained on LiDAR sensor data in order to estimate the coordinates and the orientation of a robot in an indoor environment. This work starts from a first basic model which use only a LiDAR sensor and goes on through several evolutions till a Unet-like model that has as input both LiDAR sensor and the map of the environment. All the models have been compared using a dataset of real measurements and results show that the Unet-like model is able to provide accurate position estimates with an error under the 0.2 m on average in a path. Results shows that a Deep Neural Network approach is a promising solution for robotics navigation which make use of Laser