Research And Application Of Multi-robot Obstacle Avoidance Navigation Based On Deep Reinforcement Learning

With the continuous development of artificial intelligence technology,large-scale mobile robot groups have been widely used in factory processing,logistics transportation,warehousing management,express delivery and other scenarios.In addition to requiring robots to effectively avoid obstacles and reach the destination,users usually hope that the navigation efficiency of robots can be higher and energy consumption can be lower.Commonly used technologies for robot obstacle avoidance navigation include Path Planning,Supervised Learning(SL),and Deep Reinforcement Learning(DRL).Navigation strategies based on path planning require high-quality environmental perception data and are difficult to apply in large-scale robot scenes.Navigation methods based on supervised learning require high-quality navigation datasets.Such data is difficult to collect,resulting in weak algorithm generalization ability.Navigation strategies based on deep reinforcement learning also have problems with imperfect performance in specific scenarios.For example,it is difficult to walk straight in a scene without obstacles,and the collision rate is high in a densely populated obstacle scene.To address the above problems,this paper proposes a reinforcement imitation learning navigation framework based on collision prediction(CP-RIL).Firstly,imitation learning is used to improve the navigation performance of reinforcement learning algorithms in densely populated obstacle scenarios.Then,a collision prediction algorithm based on an encoder-decoder model is designed to address the problem that learning-based end-to-end navigation algorithms have weak danger awareness.The CP-RIL framework is constructed by introducing the collision prediction algorithm based on the basis of reinforcement imitation learning.Finally,this paper implements a virtual-real combined multi-robot navigation simulation verification system for algorithm verification.The experimental results show that this method can adaptively select appropriate navigation strategies in different types of environments while completing obstacle avoidance navigation,reduce the collision rate,navigation time,and shorten the navigation distance.The main work of this paper can be summarized as follows:(1)Proposing a reinforcement imitation learning navigation method.Firstly,a navigation strategy is trained according to the reinforcement learning policy,and multiple navigation performance indicators are optimized by decomposing the multi-target reward function.Then,the method that has been validated is used to test and collect a dataset in a specific scenario.The navigation strategy is further trained according to the imitation learning method to improve the navigation performance of the reinforcement learning navigation strategy in specific scenarios.The experimental results show that the algorithm can perform autonomous obstacle avoidance in unknown static and dynamic environments,and has excellent performance in four performance indicators: speed,turning speed,distance,and time.(2)Proposing a robot collision prediction algorithm based on a long short-term memory network encoder-decoder model.To address the problem of high collision rates in densely populated obstacle scenes for learning-based navigation strategies,an encoder-decoder model is built using a long short-term memory network to predict the probability of collision of the robot in the current state based on the continuous sensor data observed by the robot and the motion commands.This collision prediction model is introduced on the basis of the reinforcement imitation learning navigation algorithm.The robot’s environment is classified based on the collision prediction results and sensor data,and traditional navigation controls are used in different environments.The experimental results show that the proposed collision prediction algorithm can effectively reduce the collision rate of robots in densely populated obstacle scenes,and the classification accuracy of robot environments is high.(3)Based on the aforementioned model algorithms and combined with actual application scenarios,a virtual-real hybrid multi-robot navigation simulation and validation system was designed and developed.The system is divided into two parts: the upper computer system and the robot control system.It has implemented functions such as order management,robot management,scheduling control,monitoring display,simulation navigation,and real-world navigation.This system successfully applied the collision prediction-based reinforcement imitation learning robot navigation proposed in this paper to actual multi-robot navigation scenarios,achieving efficient task allocation and safe navigation in multi-robot navigation scenarios,and improving production safety and efficiency for users of various types of multi-robot navigation.The research achievements of this paper provide feasible ideas and methods for the further development and application of multi-robot navigation.