Development of the Neural-Based Navigation System for a Ground-Based Mobile Robot

Abstract

In this paper, the proposed collision avoidance D* algorithm is trained, which enables the mobile robot to learn by itself from its experiences, and then fit a mathematical model by updating the parameters of a convolution neural network. The trained neural network architecture is capable of choosing an action directly based on the input images using the convolution 2D layers (CNN) architecture. A convolution neural network analyzes the exact situation using maps 30 × 30 information on its environment and the mobile robot navigates based on the situation analyzed through Deep learning. The main result of this article is a new iterative algorithm of the training set development. In the first iteration the start training set is developed, and initial learning of a neural network is made. In the next iterations trained at the previous stage neural network is used as filter for the next training sets. The filter selects the trajectories with collisions reasoned by errors of the neural network. Moreover, the numbers of convolution and fully connected layers are increased iteratively. Thus, proposed algorithm allows develop both training set and neural network architecture. Comparison of the training results for filtered and unfiltered sets is performed. A high effectiveness of the filtering is conformed. The algorithm can be used to develop the planning unit of the control system of mobile ground-based robots. The developed algorithm could be used for supervised and reinforcement learning. Modeling results of the mobile robot control system with trained neural network is presented.