基于多尺度特征融合和抓取质量评估的抓取生成方法

Grasp generation method based on multiscale features fusion and grasp quality assessment

在非结构化场景中,物体的6-Dof抓取是智能服务机器人领域的一项极具挑战性的任务。在该场景中,机器人需要应对不同大小和形状的物体以及环境噪声等因素的干扰,因此难以生成准确的抓取姿态。针对此问题本文提出一种基于多尺度特征融合和抓取质量评估的6-Dof抓取姿态生成方法。首先,提出了自适应半径查询法,解决真实环境中点云采样不均匀导致的关键点查询异常的问题;其次设计了一种将多尺度特征和抓取质量评估融合的抓取生成网络,可以生成丰富的6-Dof抓取域;最后,定义了一种抓取质量评估方法,包含力闭合分数、接触面平整度、棱边分析和质心分数,并将这些标准应用在标准数据集上生成新的抓取置信分数标签,同时将这些标准融入抓取生成网络中。实验结果表明所述的方法与当前较为先进的方法FGC-GraspNet相比平均精度提升了5.9%,单物体抓取成功率提升了5.8%,多物体场景的抓取成功率提升了1.1%。综上所述,本文所提出的方法具备有效性和可行性,在单物体场景和多物体场景中具有较好的适应性。

In unstructured environments,the 6-DoF object grasp is a highly challenging task in the field of intelligent service robotics.In such scenarios,robots need to deal with interferences from objects of different sizes and shapes,as well as environmental noise,making it difficult to generate accurate grasp poses.To address this problem,this article proposes a grasp generation method based on multi-scale features fusion and grasp quality evaluation.Firstly,an adaptive radius query method is introduced to solve the issue of key points query anomalies caused by uneven point cloud sampling in real environments.Secondly,a grasp generation network is designed to fuse multi-scale features and grasp quality assessment,which enable the generation of rich 6-DoF grasp candidates.Finally a grasp quality assessment method is defined,which includes force closure score,contact surface flatness,edge analysis,and centroid score.These criteria are applied to generate new grasp confidence score labels on a standard dataset and incorporated into the grasp generation network.Compared with the current state-of-the-art method FGC-GraspNet,the experimental results show that the described method improves the average accuracy by 5.9%,the success rate of single-object grasp by 5.8%,and the success rate of multi-object scene grasp by 1.1%.In summary,the proposed method has effectiveness and feasibility,which has good adaptability in single-object scenes and multi-object scenes.