Path Planning for Mobile Robot Considering Turnabouts on Narrow Road by Deep Q-Network

Abstract

This paper proposes a path planning method for a nonholonomic mobile robot that takes turnabouts on a narrow road. A narrow road is any space in which the robot cannot move without turning around. Conventional path planning techniques ignore turnabout points and directions determined by environmental data, which might result in collisions or deadlocks on a narrow road. The proposed method uses the Deep Q-network (DQN) to obtain a control strategy for path planning on narrow roads. In the simulation, the robot learned the optimal velocity commands that maximized the long-term reward. The reward is designed to reach a target with a smaller change in robot velocity and fewer turnabouts. The success rate and the number of turnabouts in the simulation and experiment were used to evaluate the trained model. According to simulation and environmental data, the proposed strategy enables the robot to travel on narrow roads. Additionally, these outcomes demonstrate comparable performance on a number of roadways that are not part of the learning environments, supporting the robustness of the trained model.

Keywords

• Mobile robot
• path planning
• turnabout
• reinforcement learning