Deep Reinforcement Learning with Pedestrian Trajectory Prediction Model for Service Robot Navigation in Crowded Environments

Safety and effectiveness are crucial considerations in the navigation of robots within crowded environments. Previous research in this field has predominantly focused on an omniscient approach, designing environments with exceptional obstacle avoidance capabilities. However, this approach fails to address the constraints imposed by real-world applications’ limited field of view. Additionally, constructing a robot with a complete field of view can be prohibitively expensive. To tackle these challenges, this paper presents a novel approach that combines a social spatial-temporal graph convolutional neural network (Social-STGCN) with reinforcement learning (RL) to enhance crowd avoidance navigation. This approach offers a more practical and cost-effective solution. Simulations conducted to evaluate our proposed method demonstrate its effectiveness by achieving higher success rates and lower collision rates compared to existing approaches. These results confirm the efficiency and safety of our algorithm for crowd avoidance navigation. In practical applications, we employ the Intel RealSense L515 camera to capture depth and RGB image data, enabling the utilization of the Deep SORT algorithm for pedestrian tracking and recognition. This facilitates the acquisition of accurate position and speed information. To validate the performance of our system, we conducted robot experiments in real-world scenarios, including indoor corridors and elevator lobbies. The results obtained from these experiments clearly indicate that our proposed system successfully accomplishes navigation tasks with improved safety and effectiveness. In summary, this paper introduces an effective crowd avoidance navigation algorithm. By leveraging the Social-STGCN and RL techniques, our algorithm enables robots to navigate tasks with enhanced safety and efficiency, even when their field of view is limited.