| The coming of big data era and the increasing computing capacity of computers have given rise to numerous breakthroughs in the field of deep learning like convolutional neural networks(CNNs).Since CNNs are capable of learning good feature representations from training data and achieving better performance than hand-crafted features,they are already widely used in object recognition and object detection for robots.This paper gives a detailed study on applying convolutional neural networks to robot grasp planning to give robots the capability to manipulate objects encountered in unstructured environment like warehouse or households.This paper proposed two grasp detection convolutional neural networks that are capable of online planning of grasps for unknown objects,established a complete grasp planning system and also verified the performance of the proposed system via experiments.This paper first established a complete robot grasp planning framework,and then built models for depth camera,robot grasp,and also CNNs-based grasp planning model,which laid the foundation for the study of grasp detection algorithms.Based on the established grasp planning framework and model,this paper proposed a two-stage grasp detection model named GDN,which first samples numerous candidate grasps and then use convolutional neural network to classify good grasps.This model has a novel structure that takes in images and candidate grasp angles at the same time which enables the separation between the predictions for different grasp angles,which better captures the characteristic of grasp.Transfer learning is used to apply high-performance Image Net pretrained models to the extraction of image features for grasp detection task.The established GDN model reached 79.4% validation accuracy on CMU grasp dataset.To improve the speed of grasp detection process,and skip the time-consuming candidate grasp sampling step,the paper proposed an end-to-end model named Pixel GDN,which takes RGBD image as input,and uses the extracted multi-modal features to provide a pixelwise grasp detection map.Pixel GDN outputs binary classification for each grasp angle,which avoids making the strong assumption often made by other end-to-end grasp detection model that there is only one correct grasp in each region or image.The established Pixel GDN grasp detection model reached 86.36% Top-1 grasp rectangle accuracy on the validation set of Cornell grasp dataset,and its inference speed is 27.40 fps.At last,to evaluate the performance of the proposed algorithms,this paper built a grasp simulation environment consisting of a robot,a depth camera,scene objects,and objects to be grasped using Gazebo.Then the robot grasp planning system built with ROS is deployed on the simulated depth camera and also a real depth camera to conduct grasp detection experiments.GDN has reached 86.0% and 74.1% grasp detection accuracies in simulation and real depth camera respectively,while Pixel GDN 81.3% and 71.3%.Then the paper improved the candidate grasp sampling module of GDN based on the observation during experiments,which increased the grasp detection accuracy to 89.3% on real depth camera. |
