Feasibility Evaluation of Oversize Load Transportation Using Conditional Rewarded Deep Q-Networks

It is crucial to determine feasibility when transporting oversize loads such as bridges or modular plants. In the past, swept path analysis was used to analyze the trajectory of the vehicle's movement. It is, however, a very time-consuming process. Additionally, these analysis tools do not support omnidirectional vehicles, such as the self-propelled modular transporters, which transport oversize loads. The purpose of this research is to develop a simple simulator for an omnidirectional body and a simulation-based automated feasibility evaluation system to address these problems. The DQN agent moves the vehicle and then labels the training data of the binary classifier. DQN is trained quickly and effectively using curriculum learning and a conditional reward function. Through these auto-generated labels, a binary classifier can be trained with an AUC up to 0.9606. DQN agent-based automated labeling sometimes compensates for human manual labeling errors, which is one of the most compelling findings. Furthermore, binary classifiers are about 1000 times faster than conventional swept path analysis methods. This study introduces a system for determining the transportation feasibility of oversize loads efficiently and quickly.