Application Of Deep Reinforcement Learning In Indoor Guiding Robot

At present,robots have moved from manufacturing to non-manufacturing.More and more service-oriented robots are applied in daily life.Path planning is an inevitable problem for all robots with mobile capabilities.For this problem,the traditional method has the disadvantages of long calculation time,low efficiency and poor generalization for different environments.In order to solve this problem,this thesis uses deep reinforcement learning to replace the traditional method,which does not need to describe the environment information in advance.The tedious planning process of the traditional algorithm is transformed into an end-to-end decision-making process.According to the current state information of the robot,the action order can be directly mapped by neural network,and it can make the robot have the ability of continuous navigation.The specific research contents of this thesis are as follows:Firstly,this thesis analyzes the reasons for the lack of performance of traditional algorithms,and proposes a solution based on deep reinforcement learning for indoor navigation.Secondly,the deep q-network is closely combined with the path planning problem,and the state space,action space,reward and punishment system are constructed,and the appropriate neural network structure is built.The algorithm is improved,and the traditional algorithm for single target location planning is changed to multiple continuous path planning for multiple target locations.Three kinds of indoor environments and robot models are built by using the robot operating system development platform for simulation experiments.Finally,three kinds of performance tests are carried out:(1)from low complexity to high complexity,the navigation effect of the algorithm is verified in three environments respectively.The results show that the trained models in the three scenarios can make the robot complete the continuous navigation planning for different targets with better path in a period of time.(2)The training model is applied to similar scenarios to test its generalization.It is found that the reward value converges quickly,which proves the generalization ability of the method.(3)Compared with A* algorithm and Dijkstra algorithm,the research shows that the method of deep reinforcement learning has higher real-time efficiency.