Low-cost guide dog robot navigation using Dueling DQN

Abstract. Traditional aids for the visually impaired, such as guide dogs, have limitations. They are expensive to train, require a long preparation time, and are not accessible to most people. Guide robots have the potential to address these issues. However, most existing navigation methods focus solely on the robot's trajectory, neglecting the movement of the person. As a result, the paths of the visually impaired individual and the robot may not overlap or may intersect with obstacles.To overcome these challenges, we propose an intelligent guide dog system based on a Deep Q-network (DQN). First, a kinematic model is constructed to calculate the relative positions of the guide robot and the human. Then, the reward function is designed based on this kinematic model. Finally, the Dueling DQN is trained to plan a path and predict the human's possible trajectory, ensuring collision avoidance for both the robot and the human.Our experiments demonstrated that our system performs well in both simulated environments and in real-world scenarios, such as climbing stairs. The robot can accurately find a safe path that considers the trajectories of both the robot and the human, successfully completing tasks safely.