Objects Identification And Robot Grasping In Unstructured Environment

Robot Grasping Technology plays an important role in robotic In relatively fixed industrial scenarios,using robots to perform grasping tasks has high efficiency and strong robustness However,However,objects in unstructured environment are stacked,occluded and have random placement postures with each other,which makes it difficult to recognize and complex to grasp objects.In this context,the object recognition and robot grasping operations in unstructured environment are studied.This paper includes making image datasets,training neural networks for object semantics segmentation,building grabbing models,building experimental platforms and experimental verification.The main contents are as follows:(1)The method of image data annotation is studied.The annotation specification of dataset for image semantics segmentation is analyzed.A fast image annotation method is proposed based on feature matching and Gauss mixture model.A user interface is designed,which has three modes: automatic,semi-automatic and manual.Under the condition of ensuring the accuracy of annotation,the speed of annotation is at least twice as fast as that of manual annotation.(2)The technology of object semantics segmentation based on neural network is studied.The key technologies and main processes of deep-learning are studied,and the framework of Mask R-CNN network is emphatically analyzed.Mask R-CNN network is reproduced in TensorFlow,and model parameters are adjusted by using self-made datasets to make it more suitable for unstructured environment where robot picks living things.(3)Robot grasping strategy and grasping planning are studied.A spatial parallel grabbing model for two-finger parallel gripper is proposed.The grabbing model is established by using the information of two-dimensional image space and three-dimensional point cloud space.The criteria for evaluating the quality of grabbing model are proposed,and the translation and rotation search strategies are carried out to maximize the grabbing model score.This method can quickly determine an effective grasping scheme where the shape of the object is unknown.(4)Experimental design and method verification.The software and hardware experiment platform of the robot grasping system are built,the calibration of vision system and hand-eye system are completed,and the robot grasping experiment in real environment is carried out.Experiments verify the feasibility of unstructured object recognition algorithm,grabbing pose estimation method and robot motion planning strategy.The experimental results show that the average success rates of grabbing scattered objects and stacked objects are 94.6% and 90.6%,respectively.This paper realizes object recognition in unfixed environment and grasping strategies' generation for unknown objects,which improves the intelligence and reliability of the robot grasping operation in unstructured environment.Robot grabbing system is expected to be used to solve pick-place problems instead of human in home,supermarket and other occasions.