On the training of a neural network for online path planning with offline path planning algorithms

One of the challenges in path planning for an automated vehicle is uncertainty in the operational environment of the vehicle, demanding a quick but sophisticated control of the vehicle online. To address this online path planning issue, neural networks, which can derive a heading for an operating vehicle in a given situation, have been actively studied, demonstrating their satisfactory performance. However, the study on the training path data, which specifies the desired output of a neural network and in turn influences the behavior of the neural network, has been neglected in the literature. Motivated by this fact, in this paper, we first generate different training path data sets applying two different offline path planning algorithms and evaluate the performance of a neural network as an online path planner depending on the training data under a simulation environment. We further investigate the properties of the training data that make a neural network more reliable for online path planning.