Attention-based deep reinforcement learning approach for robot navigation in dynamic environments

Abstract

The autonomous navigation of mobile robots in dynamic-static composite environments is a complex and challenging problem. The conventional obstacle recognition-path replanning approach is inadequate in terms of navigation efficiency and success rate in intricate surroundings. Our proposed End-to-End reinforcement learning autonomous navigation method from radar data to motion planning aims to enhance the autonomous navigation capability of mobile robots. We partition the LiDAR data into multiple regions based on the mobile robot coordinates and use an attention mechanism to extract semantic features of the environment. This enables the robot to perceive obstacles implicitly and adapt its motion speed and direction dynamically. Our approach was evaluated through simulation experiments in dynamic and complex scenes, and the results demonstrate that, even in the absence of map information, the method can achieve efficient and safe autonomous navigation in challenging environments.