Mobile robot navigation based on deep reinforcement learning with 2D-LiDAR sensor using stochastic approach

In recent years, there has been a significant progress in mobile robotics and their applications in different fields. Currently, mobile robots are employed for applications such as service robots for delivery, exploration, mapping, search and rescue, and warehouses. Recent advances in computing efficiency and machine learning algorithms have increased the variations of intelligent robots that can navigate autonomously using sensor data. Particularly, reinforcement learning has recently enjoyed a wide variety of success in controlling the robot motion in an unknown environment. However, most of the reinforcement learning-based navigation gets the path plan with a deterministic method, which results in some errors. Therefore, we present a navigation policy for a mobile robot equipped with a 2D range sensor based on the Proximal Policy Optimization of a stochastic approach. The tested algorithm also includes a stochastic operation, which simplifies the policy network model. We trained a differential drive robot in multiple training environments, and based on such stochastic learning, the training data accumulates faster than before. We tested our algorithm in a virtual environment and present the results of successful planning and navigation for mobile robots.