Development of a Hybrid Control Algorithm Based on Neural Networks for Mobile Autonomous Robot Navigation

A control algorithm based on adaptive neural networks is developed to control the navigation of a mobile autonomous ground-based explorer robot. A simulation was performed in the Matlab program to verify the operation of the algorithm, where the set of proximity sensors used by the robot for the detection and evasion of obstacles was simulated. The algorithm was developed using three neural network architectures: Adaline-type linear adaptive network, Perceptron neural network, and Feedforward neural network. The mathematical study of each of the adaptive neural network architectures proposed for the recognition of training patterns, corresponding to wall following patterns was carried out; their comparison was made and based on the results of the simulation, the one that showed the best output responses was selected to form part of the algorithm. This algorithm is in charged of the main tasks of obstacle avoidance and the target searching, generating as control outputs: the angle of rotation of the robot concerning its current position and its forward speed, to solve the problem of high difficulty concave obstacles encountered in the robot's path to the target. This is a hybrid algorithm, made up of a local path planning algorithm, in charge of obstacle wall fallowing and a global path planning algorithm, in charge of finding the final objective. This work focused on solving the problem of evasion of high difficulty concave obstacles, formed by curves in the form of a loop, also called dead-end traps, in which a situation of local minimum is presented. In this work different types of obstacles were simulated, being able to create almost any shape; the algorithm was tested with obstacles created from arrays of bars or lines forming simple corners, with obstacles such as U-shaped, snail-shaped, labyrinth-shaped, among other complicated shapes. The developed algorithm autonomously generates an obstacle-free navigation path, with a speed control of the robot, during the entire movement until reaching the final target. Besides, the simulation shows that the designed algorithm works adequately to solve the problem of concave obstacles, and compared with results of other mobile robot navigation techniques such as potential fields, diffuse controller-based techniques, and techniques based on rule learning (pure neural network); that in general, they present great limitations to solve the type of problem posed, difficult concave-type obstacles, remaining stuck in local minima or entering into infinite cycles with no exit, without reaching the final target; therefore, the robustness of the developed algorithm is shown.