Q-learning-based Collision-free Path Planning for Mobile Robot in Unknown Environment

摘要

With the complexity of application scenarios, higher performance requirements are imposed for the autonomous navigation ability of mobile robots. This paper proposes a Q-learning path planning method to achieve collision-free motion for the mobile robot in an unknown environment. An improved Q-learning algorithm is firstly designed by using unfixed reward function, expanded action space, and dynamic parameters in order to get an optimized and collision-free path. The convergence is reinforced by combining the gravity function in the artificial potential field algorithm and using the deep neural network instead of the Q-table. In the simulation environment, the improved Q-learning-based collision-free path planning is verified using grid map. Compared with comparison Q-learning algorithm, the path length of the improved algorithm is reduced by 17.07%, the path angle is reduced by 85.72%, and the convergence speed is shortened by 36.82%, which significantly improves the efficiency and effectiveness of path planning.