A Path Planning Algorithm Based on Deep Reinforcement Learning for Mobile Robots in Unknown Environment

The path planning is one of the hottest issues of the mobile robotics, which is to find an optimized path for the robot to reach the target location from a start position. In this research a Proximal Policy Optimization (PPO) algorithm, which is by far the most widely used and efficient Deep Reinforcement Learning (DRL) algorithm, is applied to plan a safe and collision-free path from the initial position to the target one for the mobile robot in an unknown environment where a global map of the environment is unavailable for the robot during the planning. The mobile robot can only obtain the local information of the environment through the Lidar installed on it. The path planning of the mobile robots is modeled as a Partially Observed Markov Decision Problem (POMDP), of which the state space is represented by the position of the target relative to the mobile robot and the distance information obtained by the 2D-lidar, the action space is represented by the position for the mobile robot to reach in the next step. The execution of motion controller is only conducted for the robot to implement the learned strategy rather than conducted in the whole learning episodes during the training phase, which resulted in a great improvement in the training efficiency. Computer simulations demonstrated the effectiveness of the proposed approach.