Mapless navigation based on deep reinforcement learning for mobile robots

Aiming at the problem that traditional navigation methods are dependent on map accuracy and have poor adaptability to dynamic and complex scenes, a deep reinforcement learning map-free autonomous navigation algorithm based on curriculum learning is proposed. In order to overcome the problem of learning difficulty in the case of sparse reward, a course-guided deep reinforcement learning training method based on circle of competence is proposed by drawing on the idea of curriculum learning. In addition, in order to make better use of the current collision information of the robot to assist the robot to make action decisions, the concept of collision probability is introduced, and the obstacle information currently perceived by the robot is represented in a high-level semantic form. It is encoded into the current observation of the robot as part of the input of the navigation strategy to simplify the mapping of the observation to the action and further reduce the difficulty of learning. The experimental results show that the convergence speed of the strategy is significantly accelerated after the training of the proposed course, and the success rate reaches more than 90 % in larger scenes, and the driving time is reduced by 53.5 % ∼ 73.1 %. It can provide reliable navigation for unmanned operations in unstructured unknown environments. © 2024 Northeast University. All rights reserved.