Robotic Intelligent Grasping Control Technology Based On Deep Reinforcement Learning

In the field of artificial intelligence such as deep reinforcement learning,the intelligent upgrade of robot applications is still slow.On the robot production line,some basic tasks of robots such as grasping,fixing,and handling still require the intervention of programmers.The level of automation cannot currently meet the actual production needs,and the low level of robot intelligence is hindering the development of machinery manufacturing and other industries.Therefore,to improve the robot's intelligence level by studying artificial intelligence related algorithms based on deep reinforcement learning is to meet the needs of social productivity.This paper proposes a relevant model of robot's intelligent grasping motion control,and improves the relevant algorithm in the process of training convergence.Based on the design and research of robot grasping control system,aiming at the problem that traditional robots lack the closed-loop control of grasping control action,this paper proposes a control decision-making network model and a value evaluation network model based on deep reinforcement learning.By using the strong decision-making ability of reinforcement learning,With the analysis of real-time images in the process of grasping,the robot is controlled in the current state and performs the appropriate action to complete the grab,and the simulation environment of robot system is established to reduce the training difficulty of the model.By training the control decision network,the agent interacts with the environment and updates the parameters,analyzes the system deviation caused by the difference in variance when the importance sampling is insufficient,and draws attention to the parameter update of the control decision network.It is concluded that the two probability distributions cannot be too different,and a method for coordinated updating of network parameters is proposed.An objective function to limit the update of network parameters with too large difference in probability distribution is given.The effectiveness of the method on model training results is verified in experiments.Aiming at the problem that network training data is obtained by agents interacting in the simulation environment of the robot system and thus is relevant,it cannot guarantee the independence of the data.In the process of training intelligent robots,the asynchronous training method in the simulation system is adopted.Specific algorithm implementation process.Weakened the correlation of the data,reduced the impact on the parameter convergence,and verified its effect in the experiment.In order to enhance the generalization of grasping control of intelligent robots and improve their ability to understand the environment,randomness is added to the simulation environment ofthe robot system,and a specific algorithm flow for adding randomness is given.Aiming at the problem of sparse feedback,a hierarchical reward method is proposed.Based on the above research,the robot intelligent grasp control test system is built and the robot intelligent grasp control test experiment is completed,which validating the effectiveness of the algorithm.