Obstacle Avoidance Skill Learning Algorithm For Home Service Robot Based On Improved Deep Reinforcement Learning

With the rapid development of artificial intelligence,combined with reinforcement learning technology and robot technology,research on robot operation skill learning systems with certain autonomous decision-making and learning capabilities has gradually become an important branch of robot research.As an indispensable member in the field of robots,family service robots are becoming more and more important in the fields of family services and medical assistance.Autonomous obstacle avoidance capability is the most basic and most important capability of family service robots.With the increasing complexity of family living environment,traditional family service robots have the problems of low model training efficiency and poor obstacle avoidance ability.Therefore,in view of a series of problems such as low obstacle avoidance success rate,long training time,and can only be applied in specific environments in the execution of obstacle avoidance tasks for home service robots,designing algorithms with independent decision-making and learning capabilities has become the focus of robot research..This paper studies the problem of learning obstacle avoidance skills for home service robots.Based on a comprehensive analysis of the domestic and foreign research status,it combines deep reinforcement learning algorithms and improves them.Based on this,a more efficient home service robot obstacle avoidance method is designed.The specific research contents are as follows:Aiming at the problems of long training time,over-estimated Q value,low success rate of obstacle avoidance,and can only be applied in specific environments in traditional reinforcement learning algorithms in the field of obstacle avoidance,a CDDN combining convolutional neural network and duel network architecture is proposed algorithm.First,the monocular vision sensor is used to replace the traditional laser ranging sensor.Based on the advantages of the convolutional neural network in image processing,the learned data is transferred to the experience pool of the duel network architecture.Use the duel network architecture to decompose actions into action advantages and value states,introduce an experience playback mechanism,select the next action based on the value state,and then backpropagate to update the Q value.Finally,a comparative experiment is designed to verify that the CDDN-based obstacle avoidance method solves the problems that the Q value of home service robot obstacle avoidance is too high,it can only be applied in specific environments,and the obstacle avoidance success rate is low.Aiming at the problems that the CDDN-based obstacle avoidance method has too long training time in the obstacle avoidance problem and the application in continuous motion space is still not ideal,the ACDN home service robot obstacle avoidance algorithm is proposed.First of all,the CDDN algorithm is improved in conjunction with Actor-Critic,and the policy gradient is used to control the duel network architecture in CDDN for autonomous action selection.Using the advantage of Actor's strategy update to update the entire network module,the training time of the robot is shortened,and the success rate of obstacle avoidance in the continuous action space is improved;finally,the experiment proves the deep reinforcement learning based on ACDN The performance of obstacle avoidance methods.