This paper proposes an improved Soft Actor–Critic Long Short-Term Memory (SAC-LSTM) algorithm for fast path planning of mobile robots in dynamic environments. To achieve continuous motion and better decision making by incorporating historical and current states, a long short-term memory network (LSTM) with memory was integrated into the SAC algorithm. To mitigate the memory depreciation issue caused by resetting the LSTM’s hidden states to zero during training, a burn-in training method was adopted to boost the performance. Moreover, a prioritized experience replay mechanism was implemented to enhance sampling efficiency and speed up convergence. Based on the SAC-LSTM framework, a motion model for the Turtlebot3 mobile robot was established by designing the state space, action space, reward function, and overall planning process. Three simulation experiments were conducted in obstacle-free, static obstacle, and dynamic obstacle environments using the ROS platform and Gazebo9 software. The results were compared with the SAC algorithm. In all scenarios, the SAC-LSTM algorithm demonstrated a faster convergence rate and a higher path planning success rate, registering a significant 10.5 percentage point improvement in the success rate of reaching the target point in the dynamic obstacle environment. Additionally, the time taken for path planning was shorter, and the planned paths were more concise.

Although numerous effective Simultaneous Localization and Mapping (SLAM) systems have been developed, complex dynamic environments continue to present challenges, such as managing moving objects and enabling robots to comprehend environments. This paper focuses on a visual SLAM method specifically designed for complex dynamic environments. Our approach proposes a dynamic feature removal module based on the tight coupling of instance segmentation and multi-view geometric constraints (TSG). This method seamlessly integrates semantic information with geometric constraint data, using the fundamental matrix as a connecting element. In particular, instance segmentation is performed on frames to eliminate all dynamic and potentially dynamic features, retaining only reliable static features for sequential feature matching and acquiring a dependable fundamental matrix. Subsequently, based on this matrix, true dynamic features are identified and removed by capitalizing on multi-view geometry constraints while preserving reliable static features for further tracking and mapping. An instance-level semantic map of the global scenario is constructed to enhance the perception and understanding of complex dynamic environments. The proposed method is assessed on TUM datasets and in real-world scenarios, demonstrating that TSG-SLAM exhibits superior performance in detecting and eliminating dynamic feature points and obtains good localization accuracy in dynamic environments.