Research On Robot Pushing And Grasping Operation Skills Based On Deep Reinforcement Learning

With the rapid development of artificial intelligence,deep reinforcement learning algorithms are widely used in the field of robotics to solve the problem of learning robot operating skills.Pushing and grasping skills are the most basic and main skills of home service robots.In view of the high cost of training and low action efficiency in the learning stage of home service robot target promotion and grasping operation skills,this article combines deep reinforcement learning algorithms and improves them,based on this design more efficient home service robot goal promotion and grasping operations Skill learning algorithms to improve the perception and autonomous decision-making capabilities of home service robots in complex environments.The specific research content is as follows.First of all,in order to solve the problems of low learning efficiency and low success rate when the robot pushes and grasps the target object in a complex environment,it proposes a deep Q-learning robot operation skill learning algorithm based on visual attention.First of all,the attention mechanism is introduced into the visual network to improve the perception of target objects in the work area,fully extract the features of the work space,and generate a target functional map.Second,the robot uses the self-supervised training deep Q-learning fitting to calculate the Q value of the discrete actions of “pushing”and “grasping”,and chooses to execute the next action.Finally,a robot simulation environment was established on the V-rep simulation platform,and a comparative experiment was designed to verify the effectiveness of the deep Q learning algorithm based on visual attention to solve the problem of robot pushing and grasping operation skills learning in a complex environment.Then,in order to solve the problem of unsatisfactory action coordination between the robot pushing and grasping operation skill learning algorithm based on visual attentionbased deep Q learning,a deep reinforcement learning algorithm DQAC(Deep Q Actor Critic)based on the Actor-Critic framework was proposed.First of all,the entire algorithm framework is modularly designed,which is mainly divided into a visual mechanism module and an action mechanism module.The visual mechanism module fits the Q value to calculate the position and motion angle parameters of the robots actions.The action mechanism module adopts the Actor-Critic idea.To explore an optimal action selection strategy.Secondly,use dual experience pools to store image data and action data separately,and perform batch training on training data to improve the utilization of training data.Finally,experiments verify the efficiency of the DQAC algorithm in solving the problem of robot pushing and grasping operation skills learning.