Dynamic obstacle avoidance and evaluation base on neural network

Abstract

The aim of this dissertation is to address the issue of dynamic obstacle avoidance in robotics. By combining genetic algorithms and neural network technology, a novel dynamic obstacle avoidance control system is developed. The dissertation introduces the Neuro Evolution of Augmenting Topologies (NEAT) neural network as the controller for the dynamic obstacle avoidance system, enhancing both its avoidance effectiveness and generalization capability. Furthermore, the dissertation leverages artificial potential fields (APF) and a designed global path evaluation function to construct a training dataset, utilizing a multi-input Multi-Layer Perceptron (MIMLP) neural network for data fitting. Through multiple training and evaluation iterations in a simulated environment, the results demonstrate that the designed dynamic obstacle avoidance system successfully converges in random environments and exhibits superior avoidance performance and generalization ability. This performance surpasses that of existing traditional obstacle avoidance algorithms with fixed param across various environments.