Robot Visual Navigation Algorithm Based On Deep Reinforcement Learning

With the growing development of computer vision technology,navigation technology has been gradually applied in fields such as autonomous driving and human-computer interaction,and has received close attention from academia and industry.Traditional robot navigation technology builds maps first and then plan paths,which cannot meet various needs in real-time scenarios.The visual navigation technology based on deep reinforcement learning can adaptively identify target features according to different scenes,and achieve better navigation effect.Therefore,the study of visual navigation technology based on deep reinforcement learning has extremely important theoretical significance and application value.Aiming at the problem that the navigation performance of the agent is greatly weakened due to the change of the navigation scene in the target-driven visual navigation system,a deep reinforcement learning visual navigation model integrating long short-term memory network(LSTM)is proposed.The model needs to input RGB images of the current state and target state to realize visual navigation.On the basis of improving the original target-driven visual navigation model,based on historical state information,combined with LSTM and Universal Successor Representation(USR),it makes decisions on future actions.Experiments are carried out in the AI2-THOR simulation environment.The results show that the agent trained by the model proposed in this paper has excellent navigation performance.Compared with other models,the average path length is reduced by about 6%,the average collision rate is reduced by 40%,and the model converges quickly.faster.Aiming at the problem of reduced accuracy due to navigational scene changes in targetdriven visual navigation systems,a visual navigation model that fuses attention mechanism and next expected observations(Neo)is proposed.On the basis of the original Neo,the CSPRes Ne St50 network is added to improve the recognition accuracy of the main features of the current observation image by the agent,and the loss calculation method is improved,thereby improving the accuracy of the navigation model in different scenarios.In the AVD dataset and AI2-THOR scenarios,the training and testing are performed respectively.The results show that the model proposed in this paper has excellent performance.Compared with other models,the average SR is improved by about 3%,and the average SPL is improved by about 6%.