| Grasping is one of the most important means for a robot to interact with the environment and can be applied to a wide range of scenarios,such as pick-and-place,bin-picking,conveyor tracking and so on.As grasping scenes become increasingly complex,for example,from a simple single-object scene to a multi-object stacking scene,the demand for intelligent grasping is also increasing.In the scene of object stacking,obtaining the relationship and the manipulation order of objects is a necessary way to realize higher-level interaction with the environment.However,many existing methods focus on individual object features and ignore contextual information,leading to great challenges in efficiently reasoning manipulation relationship.More importantly,when the target object is completely blocked or there is no suitable grasping position around the target object,the robot still needs to infer the intelligent action sequence,and the cooperation of grasping and pushing actions based on reinforcement learning is an effective method to solve this problem.However,many existing methods rely too much on model design,or there are redundant operation actions.To solve the above problems,this thesis uses graph convolution network to collect the context information between objects,which can directly output the relationship and manipulation order of objects;In addition,this thesis also combines graph convolution network with reinforcement learning to better realize robot grasping and pushing cooperation;We also build the robotic autonomous grasping system to realize the robotic autonomous decision-making and grasping manipulation.The main research work of this thesis is as follows:1.Research on graph-based visual manipulation relationship reasoning network for robot graspingMost of the traditional visual manipulation relational reasoning rely on convolution layer and full connection layer,and they realize relational reasoning by traversing all objects.However,with the increase of the number of objects,redundant information is also increasing significantly and the efficiency of relational reasoning is reduced.In this thesis,a graph-based visual manipulation relationship reasoning network is proposed,which focuses on representing the information in the image as objects and the relationship between them,and inferring the object relationship and manipulation order directly through graph convolution network.Our model first extracts features and detects objects from RGB images,and then uses graph convolution network to collect context information between objects.In addition,a relational filtering network is established to reduce the object pairs before relational reasoning and improve the efficiency of relational reasoning.Experiments on visual manipulation relation dataset(VMRD)show that our model significantly outperforms previous methods on reasoning object relationships in object-stacking scenes.2.Research on graph-based deep reinforcement learning network for robot graspingWhen exploring completely covered objects,previous methods can not learn additional knowledge from the training process to adapt to different scenes,and it is unreasonable to use the learned push prediction value when the target object is visible to estimate the operation position when the target object is completely occluded.This thesis proposes a graph-based deep reinforcement learning model.The model first extracts the state feature X,and then uses different graph based Q networks to estimate the Q value according to whether the target object is found or not.Graph-based Q network model includes encoder,graph reasoning and decoder.The encoder is used to integrate the state features so that the features of one region include the features of other regions.Graph reasoning captures the internal relationship of features between different regions through graph convolution network.The decoder maps the transformed features after reasoning to the original state features.We also extended the test dataset for grasping and pushing cooperation,including 20 different scenarios,in which the number of stacked objects ranges from 2 to 5.In the experiment exploring the target object task,our model can not only effectively explore the completely occluded object,but also greatly improve the performance of the task of grasping and promoting cooperation.3.Construction of robot grasping systemIn order to verify the applicability of our model in the real environment,we constructed an intelligent grasping system,which is mainly based on AUBO-i5 manipulator and Intel D435 i depth camera.We build the system in many aspects,such as hardware system,software system and camera calibration,and finally realize the communication and control between the algorithm module under Linux system and the manipulator in the real environment.We verify the proposed graph-based visual manipulation relationship model and graph-based deep reinforcement learning model respectively in the real robot.The experimental results show that it is very necessary to consider orderly grasping in object-stacking scenes,and the visual manipulation relationship of our model in the real environment can still achieve good results,which reflects the practicability and generalization of our model.In addition,in the grasping experiment of occluded objects,our graph-based Q network model has higher rate for successful grasp with less motions,which reflects that our model can achieve more efficient grasping and pushing cooperation.This research work solves the problem that robots can make intelligent and sequential grasping decisions in cluttered scenes.For the visual manipulation relationship reasoning in object-stacking scenes,from the perspective of graph structure,the characteristics of the image are closely combined with the attributes of the graph to improve the accuracy of relationship reasoning.For grasping occluded objects,we use graph convolution network to realize Q network in reinforcement learning,which improves the adaptability of the model to different scenes and the performance of grasping and pushing cooperation.To sum up,our research is of great significance to the intelligent grasping of robot. |
