End-to-End learning-based navigation of autonoumous mobile robot

Abstract

In this work we present an end-to-end learning approach that is able to perform target- oriented navigation and collision avoidance using Deep Neural Network. This approach can be defined as learning a model that maps sensory inputs, such as raw 2D-laser range findings and a target position, to navigation actions for controlling the mobile robot such as steering commands. Compared to the traditional autonomous navigation systems, which often require perception, localization, mapping, and path planning, the end-to-end learning approach offers a more efficient method. which utilize large set of expert navigation demonstrations to learn the desired navigation policy.

The end-to-end learning approach has gained considerable interests in autonomous navigation in academic and industrial fields. Researches have already used different artificial neural networks to predict steering commands. However, most of the existing end-to-end methods are used for lane keeping for self-driving cars. therefore, we propose an end-to-end navigation model for mobile robots that is based on a Convolutional Neural Network (CNN). The network was trained using expert demonstration data which was generated in virtual simulation environments. The learned model was test in real time simulation and gave an acceptable result, however, it suffered when it encounters situations that requires hard maneuvers. Therefore, in order to overcome some of these difficulties, we proposed an improved model which incorporates the temporal information in the prediction process using the Long Short-Term Memory (LSTM) network. basically, this model aims to include the motion history of the robot in the steering prediction model. The improved model showed its ability to predict steering commands with high performance compared to the expert operator. However, this model imposed some limitations which will be further discussed in this remaining parts of this thesis.