Convolutional Neural Network Based Unmanned Ground Vehicle Control via Deep Reinforcement Learning

In order to reduce the cost of human resources and material resources and improve the power line inspection efficiency, unmanned ground vehicle (UGV), which utilizes the modern artificial intelligence such as deep learning and reinforcement learning, is commonly introduced to replace of human to inspect power lines in the grid system. This paper provides a deep Q network (DQN) and convolutional neural network (CNN) based end-to-end control model to drive UGV to inspect automatically, and meanwhile to avoid obstacles. Specifically, we utilize the preprocessed grayscale image as the input of the CNN, and output the final Q value. This model simulates human learning behavior by interaction between UGV and the environment. Through repeated self-learning and reward value increasing in a simulation environment, the UGV successfully reaches the target position in a shortest time and meanwhile avoiding a variety of obstacles.