Research On Multi-Degrees-of-Freedom Grasping Detection Method For Robot In Complex Scenes

Grasping is a basic skill that is essential for robots in various fields such as industry,agriculture,and firefighting,as it can improve the efficiency of their work.Currently,the planar grasping detection algorithms have issues such as low accuracy in complex scenes,and weak perception of global information.Six degrees of freedom grasping algorithm suffers from weak generalization ability due to the discrepancy between offline training data and real test scenarios,less consideration of multi-scale information of seed points,and crude pooling of seed points neighbourhood features.In response to these issues,this study is based on deep learning and focuses on enhancing the ability to extract global and local information,and adaptive feature capabilities.The main research content and innovative points of this paper are as follows:(1)Aiming at the problem that traditional grasping detection methods are not strong enough to perceive global information,this paper proposes a planar grasping detection method based on Multi-head Self-Attention Convolution and Adaptive Feature Fusion Network(MSACAFF-Net).This method first realizes the local feature extraction in the window by building a window multi-head self-attention convolution model,and builds a large kernel convolution module to model the global information,realizes the flow of information between windows,and improves the network’s ability to perceive the global and local;Then through the Adaptive Feature Fusion Network(AFF-Net),the features are fused according to the importance of the grasping detection,the features related to the grasping detection are enhanced,and the correlation between the features and the grasping parameter space is improved;Finally high-precision pixel-level plane grasping pose parameters are output through three parallel branches.(2)The Hybrid Metric Evaluation Model(HMEM)is designed to address the problem that the offline data obtained by the existing 6-DoF grasping inspection method using force closure grasping quality metric differs greatly from the real test scenario.The hybrid metric evaluation model adds the flatness metric and the contact area metric,which reflect the local geometric characteristics of the object,to the force closure metric to narrow the gap between the offline data and the real scene and improve the generalization ability.The Multi-scale Perceive and Adaptive Feature Aggregation Network(MPAFA-Net)is designed to address the problem that the network lacks the perception of multi-scale information of the seed points and the maximum pooling tends to miss important features.The network obtains the neighborhood information of each grasping seed point through different scaled perceptual regions,and uses adaptive pooling and self-attention to autonomously select and aggregate neighborhood features to improve the prediction accuracy of the grasping parameters of the seed points.(3)In this paper,a robot grasping system is built,including visual perception,grasping detection,robot control and other functions,to verify the generalization performance of the grasping detection algorithm in real scenarios.Around the robot grasping system,the mobile platform is added to build a human-robot mutual assistance system based on functional grasping,containing functional detection,robot grasping,human-robot interaction,trajectory planning and other functions to achieve the purpose of human-robot mutual assistance in work scenarios.In this paper,the proposed method is tested on the Cornell dataset,PLGP dataset,the GraspNet-1 Billion dataset,and the real scene robot grasping system,and compared with the advanced algorithms in recent years,it fully proves the advanced nature and universality ofthe method.