基于PointNet++的机器人抓取姿态估计

Robot Grasping Attitude Estimation Based on PointNet++

为解决在无约束、部分遮挡的场景下对部分遮挡的物体生成可靠抓取姿态的问题,基于PointNet++网络改进了一种抓取姿态估计算法,该算法可直接从目标点云中生成二指夹具的抓取姿态。由于该算法降低了抓取姿态的维度,将抓取的7自由度问题转变成4自由度问题处理,从而简化学习的过程加快了学习速度。实验结果表明:该算法在无约束、部分遮挡的场景中,能够生成有效的抓取姿态,且较Contact-GraspNet算法成功抓取率提升了约12%,能够应用于家用机器人的抓取任务。

In order to solve the problem of generating reliable grasp attitude for partially occluded objects in unconstrained and partially occluded scenes,this paper proposed a grasping attitude estimation algorithm based on PointNet++network.The algorithm can directly generate two finger gripper grasp attitude from the target point cloud.Because the algorithm reduced the dimension of grasping attitude,it transformed the problem of 7 degrees of freedom into the problem of 4 degrees of freedom,accelerating the rate of learning by simplifying the process.Experimental results show that the algorithm can generate effective grab attitude in unconstrained and partially occluded scenes,and improves the grabbing rate prediction by about 12%compared with Contact-GraspNet,which can be used for grabbing tasks with home robots.