Robotic Autonomous Grasping Strategy And System For Mixed Stacked MULTI-CLASS Targets

In industrial production,research on workpiece gripping strategies in mixed stacked environments has been a hot topic in the field of robotic gripping.With the shortage of labor and the release of Industry 5.0 concept in recent years,the application area of manipulators is expanding,however,with the increasing complexity of the grasping environment,this has put forward higher requirements on the gripping strategy of manipulators.In this paper,we combine the more popular artificial intelligence theories and introduce reinforcement learning methods into robotic control,combining the improved deep reinforcement learning method Soft ActorCritic(hereinafter referred to as SAC)with FCN networks for the first time,and using the strategy of pushing and grasping cooperative actions,which overcomes the traditional grasping strategy planning time,low accuracy,and poor generalization in the mixed stack environment,etc.It also has higher grasping success rate and action efficiency compared to traditional reinforcement learning strategies that perform discrete actions.The main work and conclusions of this paper are as follows.Firstly,this paper illustrates the traditional manipulator grasping strategy in different scenes and analyzes the principle of manipulator grasping,based on which,the SAC algorithm based on the maximum entropy theory and the FCN network are analyzed and demonstrated in detail to verify the feasibility of the manipulator control strategy based on the SAC algorithm in solving the workpiece grasping in the overlapping environment.Secondly,the model in this paper uses RGB-D information as the visual input,segments the objects in the overlapping scene using FCN network,and jointly trains two full convolutional neural networks with modelfree SAC algorithm for the grasping and pushing actions of the manipulator,respectively,and the manipulator obtains the best action model by continuously interacting with the overlapping environment.To address the problems of robustness and convergence in the training of action models,the algorithm is improved by adding a parameter penalty term to the objective function,introducing the idea of truncated double-Q learning and incorporating a preferred experience replay mechanism(PER).Finally,the proposed SAC algorithm is experimented and analyzed for grasping control in different scenarios,using Coppelia Sim software as the simulation platform,using the Panda manipulator as the action object,modeling the overlapping environment accurately,designing the state space,action space,reward function and other related parameters,and finally conducting experimental verification.The test experiments are divided into three groups: the experiments on grasping in the mixed stack environment,which verify the necessity of SAC algorithm to perform workpiece grasping by pushing and grasping actions with each other;the experiments on the effect of different number of workpieces on grasping performance,which verify that the manipulator needs more actions to complete grasping as the degree of mixed stack increases;the experiments on the generalization of the model,which verify that the deep reinforcement learning algorithm,compared with other traditional methods,is more effective when the workpieces in the scene The advantage of the model does not need to be retrained when the artifacts in the scene are replaced with objects whose physical parameters do not differ much.Experiments on grasping in a real environment verify the feasibility of applying the method to practical engineering tasks.The results show that the strategy of using the improved SAC algorithm to control the manipulator with a combination of grasping and pushing actions can excel in the task of grasping workpieces in complex overlapping environments,sacrificing action efficiency in exchange for a higher grasping success rate.Compared with reinforcement learning algorithms and traditional control algorithms that perform discrete actions,the success rate is higher,the action efficiency is higher,and the generalization is better.The method proposed in this paper effectively solves the drawback that traditional strategies are difficult to complete the workpiece grasping effectively in the mixed stack environment,and has good practical value.