| The ability to grasp objects is one of the most important and fundamental capabilities of intelligent robots.With the great progress made by deep learning algorithms in visual perception,many learning methods have been applied to grasping techniques,and the biggest advantage of these deep learning methods is the higher robustness in grasping objects unknown to the system.However,the placement of objects in the real environment is often disorderly or complex stack,the robot arm in this environment will often grasp poor real-time grasping,can not find a reasonable grasping position,grasping the order of confusion,the robot arm jaws and the object to be grasped collision and other problems,so how to more efficiently grasp unknown objects in the unstructured environment of multi-object clutter stack is still a challenging problem.To solve the above problem,this paper proposes a new robotic arm grasping algorithm based on priority strategy to achieve real-time collision-free grasping by predicting the topmost object in a cluttered stacked scene from visual information.Specifically,firstly,this paper proposes a new lightweight generative convolutional neural network based on grasping priority to obtain a mapping model of image information to grasping parameters,in which long-range dependencies are obtained by modeling spatial context through self-calibration and self-attention mechanisms to achieve a more discriminative feature representation and an expanded perceptual field from visual input capture.Secondly,a grasp priority policy is designed to solve the problem of real-time collision-free grasping in a multi-object cluttered stacking scene,in which the network identifies the objects closer to the top layer in the cluttered stacking scene and gives higher grasp priority to the top layer objects to achieve collision-free grasping in the stacking situation.In addition,this paper designs a priority-based position loss function to suppress the feature expression of bottom objects in multi-object clutter stacking scenes,so that the attention of the network is more focused on the top objects,weakening the noise interference of the images,reducing the overall computation of the system,and obtaining a higher real-time performance while improving the grasping performance.Finally,a robotic arm grasping platform is built in both simulation and real environment to verify the effectiveness of the algorithm in this paper.For the performance testing part,this paper conducts a large number of comparison experiments and ablation experiments for performance evaluation,detailing the superiority and generality of this paper’s algorithm from both qualitative and quantitative aspects,and obtaining the state-of-the-art crawling performance on REGRAD standard grasping dataset.In the ablation experiments,this paper demonstrates that the proposed grasping priority strategy and priority-based position loss function have a relatively large impact on the grasping performance,where the grasping priority strategy can improve the grasping success rate of the whole system by 13.4%.In order to verify the grasping performance in real scenes,this paper conducts several sets of grasping experiments on a real robotic arm grasping platform,which proves that the algorithm of this paper shows superior grasping performance in fewer objects and simpler stacking relationships,and shows a higher degree of generalization ability in scenes with more objects and complex stacking relationships. |
