Memory-based soft actor–critic with prioritized experience replay for autonomous navigation

Due to random sampling and the unpredictability of moving obstacles, it remains challenging for mobile robots to effectively learn navigation policies and accomplish obstacle avoidance safely. Overcoming such challenges can reduce the time cost required for navigation model training and validation, improving the safety and credibility of autonomous navigation in medical service and industrial patrol. This article proposes an improved soft actor–critic model to enhance the autonomous navigation performance of robots. We first introduce a prioritized experience replay method to reduce the randomness of sampling. The performance of the navigation policy can be enhanced by prioritizing the learning of high-value experiences. Moreover, we also design a network with long short-term memory abilities to store historical environmental information. In this way, temporal characteristics of obstacle motion can be obtained to optimize obstacle avoidance policy. Experimental results in simulation and real-world show that the proposed model significantly improves learning speed, success rate, and trajectory smoothness while exhibiting excellent obstacle avoidance performance in dynamic environments.