| Robot navigation tasks are the basis of various mobile robot tasks,which require the robot to move from the starting position to the target position as quickly as possible while avoiding collisions with other objects in the scene.With the rapid development of robotics technology,mobile robots are increasingly deployed in complex scenarios with limited communication networks and diverse obstacles.Network limitations(such as limited communication bandwidth and range,wireless network interference,etc.)make it difficult for robots to exchange information seamlessly,and the diversity of obstacles(such as irregularly shaped static objects,moving pedestrians,etc.)increases the difficulty of collision avoidance.With the background of complex scenes with limited communication and diversified obstacles,this project conducts related research on robot collaborative navigation tasks in response to the above challenges.In this dissertation,the cloud robot platform is used to expand the collaborative capabilities of multi-robot systems.The multi-agent reinforcement learning algorithm is used to enable robots to learn the ability of selective communication to improve the efficiency of collaborative navigation.The auxiliary task of pedestrian trajectory prediction is introduced to improve the network model’s representation ability of human motions,thereby improving the robot’s collaborative navigation performance in dynamic pedestrian scenes.Specifically,the research content of this article includes the following three parts:(1)This dissertation proposes a multi-robot cooperative offloading method with Quality of Service(Qo S)in a communication-constrained scenario.Aiming at the limited communication bottleneck problem in the multi-robot system,this project improves the existing cloud robot platform that supports multi-robot cooperative offloading,and proposes a corresponding Qo S guarantee mechanism at two levels,which are cooperative offloading decision-making and cloud platform container management.The platform ensures the Qo S when the cloud platform assists in the execution of multi-robot tasks in complex scenarios,so as to improve the efficiency of collaborative navigation.(2)This dissertation proposes a multi-robot collaborative navigation reinforcement learning algorithm that supports selective communication.Aiming at the communication limitation of mobile robots,this project utilizes the multi-agent reinforcement learning method augmented with communication,and proposes a multi-robot navigation framework that combines agent-level information and sensor-level information.This method is not only suitable for complex scenes containing multiple obstacles,but also allows each robot in the multi-robot system to comprehensively consider the surrounding environment and autonomously select the most valuable communication message.While satisfying the limited communication bandwidth,the collaboration efficiency between multiple robots is maximized,and a collaborative navigation strategy is generated.(3)This dissertation proposes a collaborative navigation reinforcement learning method for complex scenes containing dynamic pedestrians.Aiming at complex scenes containing multiple dynamic pedestrians,this project uses spatio-temporal graph attention neural network to extract the cooperative attributes and potential features between robots and dynamic pedestrians.At the same time,a selfsupervised learning method based on pedestrian trajectory prediction is introduced as an auxiliary task to assist the training process of the robot navigation strategy,so that the neural network model has a more adequate understanding and representation of the humans motion,and generates more efficient collaborative navigation strategy.Based on the above-mentioned implementation schemes and mechanisms,this dissertation conducts experiments in a variety of robot collaborative navigation scenarios to prove the effectiveness and advantages of the proposed methods. |
