Research On Robotic Grasping Based On Deep Reinforcement Learning

With the development of deep learning and computer vision,the application of intelligent vision robots in daily life in industry is gradually increasing.Many algorithms are based on supervised learning and reinforcement learning for robotic arm grasping tasks,but their ability to understand complex environments and corresponding grasping capabilities is limited.Most of the existing industrial-level applications focus on targeting a specific task.Optimization.In order to improve the efficiency and accuracy of robotic arm grasping tasks,more and more deep learning methods have been applied to robotic arm grasping scenarios in recent years,among which data augmentation is an essential technical means—generating new data To improve the model’s generalization ability and then improve the perception ability of the visual model.However,most current algorithms only consider a fixed viewing angle,and their performance is limited for scenes with irregular shapes or multiple interlaced objects.In contrast,multi-view algorithms can synthesize visual information from different angles to provide more comprehensive and accurate scene perception.However,the current multi-view algorithms usually adopt a single-step decision-making method.Most assume that the agent can obtain the global information of the environment in the initial state,which has a considerable gap with the actual application environment.Therefore,it is necessary to design a multi-view algorithm more suitable for the actual application environment to improve the efficiency and accuracy of the robotic arm grasping task.In response to the above problems,this article:1.Proposed a data augmentation strategy search algorithm based on cyclic neural network controller in the robotic arm grasping scene.Adding preprocessing not only improves the data validity and generalization of the robotic arm grasping model,but also improves the training efficiency of the model.2.Models the multi-view robotic arm grasping problem as a sequence decision problem,and proposes a sequential decision-making multi-view grasping algorithm based on Q-learning,which removes the strong assumption that the agent has global information in the existing multi-view grasping task,and The use of sequential decision processes to model multi-step actions is considered for the first time,and reinforcement learning methods are used to solve multi-view selection problems.Constructing a multi-view grasping dataset of responses based on multi-view problems and conducting experiments on this basis,the experiments show that our algorithm has advantages in grasping success rate and computational efficiency.3.In view of the problems of low data efficiency and deviation in policy optimization in the proposed algorithms for multi-view capture,we will introduce comparative unsupervised representation learning to decouple representation learning and policy optimization,thereby improving sampling efficiency,and combining double-Q learning The mechanism incorporates decoupled policy optimization to address overestimation issues.The efficient multi-view grabbing algorithm is validated in experiments based on the multi-view GraspNet dataset.Compared with previous methods,efficient multi-view grasping shows higher grasping accuracy and efficiency.