基于多物理特征的抓取检测方法

Grasping Detection Method Based on Multiple Physical Features

为解决在无约束、非结构化的多物体场景下对物体生成可靠抓取姿态的问题,提高机械臂在多场景下抓取检测的准确率,我们提出了使用场景的多物理特征进行抓取检测的评估方式,基于PointNet 网络提出了一种前置–后置联合网络抓取姿态估计算法,该算法利用多物理特征从目标点云中生成评估得到的置信度分数对场景进行进一步抓取检测得到可靠的抓取姿态。由于该抓取检测方法比以往的抓取检测评估更加细致,从而可以更好地从场景中得到精确的抓取姿态。实验结果表明:该算法在复杂的多物体场景中,能够生成有效的抓取姿态,且较同类型算法抓取成功率有所提升,能够应用于工业机器人的抓取任务。

In order to solve the problem of generating a reliable grasping attitude for objects in unconstrained and unstructured multi-object scenes and improve the accuracy of grasping detection of mechanical arm in multiple scenarios, we put forward an evaluation method of grasping evaluation method using multiple physical features of scenes. Based on PointNet network, we put forward a front-post joint network grasping attitude estimation algorithm. The algorithm uses multiple physical features to generate the confidence score obtained from the target point cloud to further capture and detect the scene to get a reliable grasping attitude. Because this grasping detection method is more detailed than the previous grasping detection evaluation, it can get an accurate grasp posture from the scene. The experimental results show that the algorithm can generate an effective grasping attitude in complex multi-object scenes, and the grasping success rate is improved compared with the same algorithm, which can be applied to the grasping task of industrial robots.