Research On Robot Navigation Algorithms In Dense Crowds Environment

With the development of artificial intelligence technology,robots are more and more closely related to human life.Robots have gradually become a good assistant in human production and life.Robots can not only replace human to do complex work,but also improve the quality of human life in life,ranging from sweeping cooking to automobile production.However,most of the current robots are not ideal smart robots,they can only do simple work in a simple environment.In complex environment,especially in the complex environment where people and robots coexist,how to achieve the autonomy of the robot to cope with the complex environment where people and robots coexist has become an urgent problem to be solved.Therefore,this thesis takes mobile robot as the research object,and does Algorithm Research on mobile robot navigation in dense crowds environment.The main research contents are as follows:(1)Existing navigation algorithms based on deep reinforcement learning do not process time series information between multiple frames of data,and do not consider the future motion state of pedestrians.This will result in the shortsighted problem of the algorithm’s planning and decision-making ability,and can not consider the optimum navigation from the global perspective of the navigation process,which will lead to the robot detouring in the population and even navigation failure.A navigation algorithm design algorithm based on partial observable reinforcement learning is presented.The navigation algorithm design algorithm based on partial observable reinforcement learning uses partial observable Markov decision process model instead of traditional Markov decision process model in deep reinforcement learning.The interaction between adjacent frames of observation data is incorporated into the reinforcement learning network.The new algorithm overcomes the shortsighted problem caused by the existing robot navigation algorithm based on reinforcement learning considering only the current frame.(2)Aiming at the problems that the navigation algorithm based on model-free reinforcement learning needs a lot of training data,the sample sampling rate is low,the generalization ability is poor,and the reinforcement learning model needs to be retrained when the use environment changes,a navigation algorithm design based on artificial potential field and model-based reinforcement learning is proposed.In the navigation algorithm design algorithm based on artificial potential field and modelbased reinforcement learning,this paper separates the training of environment model from the reinforcement learning control strategy,which can train the environment model independently,realizes the effect that the environment model can be changed even for different environments,and enhances the generalization ability of the algorithm.In the navigation algorithm based on model free reinforcement learning,this paper is based on model predictive control algorithm,takes the environment model based on neural network as the model prediction module of model predictive control,and takes the path planning algorithm based on artificial potential field as the model planning module of model predictive control,so that the three algorithms are combined to form a new hybrid navigation algorithm.Compared with the robot navigation algorithm based on model-free reinforcement learning,this algorithm does not need to retrain the reinforcement learning model after the change of the navigation environment,and overcomes the problems of low sample rate and poor generalization ability of the existing robot navigation algorithm based on model-free reinforcement learning.