An Application of Evolutionary Neural Networks for Mobile Robots Navigation and Dynamic Obstacles Avoidance

This paper explores an application based on evolutionary computation, in order to train a neural network for reactive navigation. The objective of the mobile robots navigation task is to reach an arbitrary destination point, without colliding with static or dynamic obstacles. The considered environment is an autonomous factory, where multiple robots (all with the same control system - the proposed neural network trained by evolutionary procedures) navigate to transfer parts between processing stations. This environment is considered to be stochastic and dynamic, because robot's paths between workstations are intertwined. The neural networks training was carried out in an evolutionary way using PSO (particle swarm optimization) algorithm in a series of simulation training maps, in order to maximize the ability to cope with different situations and complexities of environments. Finally, the validation was performed on a simulated environment (map) similar to a real factory. The obtained neural network was able to successfully navigate several robots, without collisions, always reaching the target positions.