An Efficient Planning Method for Autonomous Navigation of a Wheeled-Robot based on Deep Reinforcement Learning

In this paper, a planning method to autonomously navigate a wheeled-robot based on deep reinforcement learning is proposed. The planning algorithm aims at reaching a predetermined target configuration while avoiding obstacles. It is assumed that the robot chassis is differential-drive and it is equipped with multiple distance sensors to perceive the environment and Global Positioning System (GPS) for localization. A Deterministic Actor-Critic model is designed to navigate the robot in an environment including multiple obstacles. Since navigation is a multi-objective optimization problem, a novel reward function considering different contradicted criteria including reaching the target and avoiding the obstacles is proposed in this paper. To improve the performance and convergence of the learning method, a gradual learning method is applied to first learn target reaching and next obstacle avoidance. The effect of the proposed reward function is compared with some other reward functions to show its effectiveness.