基于模仿学习的双曲率曲面零件复合材料织物机器人铺放

Robotic Layup of Woven Composite on Double Curvature Surface Parts Based on Imitation Learning

  • 摘要: 针对双曲率曲面零件的复合材料织物铺放,手工铺放效率低、质量均一性差,机器人铺放的相关研究未能准确地描述手工铺放技能。为此,本文提出基于模仿学习的铺放技能采集、描述与重现的相关方法。首先,利用拖动示教获取织物铺放的轨迹信息,以压力阈值为分割依据进行有监督的轨迹分割,再采用高斯混合模型(Gaussian mixture model,GMM)与高斯混合回归(Gaussian mixture regression,GMR)学习得到铺放轨迹概率模型。在此基础上,以示教姿态和曲面法向为参考,计算末端执行器的法向姿态,再次进行示教获得法向压力信息,由GMM/GMR学习得到铺放压力概率模型;然后,针对接触状态建立过程,在力/位混合控制框架下,提出预控制量退出机制,实现接触状态建立过程中无冲击的控制器切换;最后,基于UR5e机械臂、ATI六维力/力矩传感器和ROS(robot operating system)搭建试验平台,进行试验。结果表明:所提出的技能采集、描述方法可实现对人工铺放经验的准确提取与表达,采用预控制量退出机制,接触状态建立过程的最大压力误差为1.8 N,小于曲面模具所有16条轨迹的铺放压力均方根误差2.0 N,完成的铺层无肉眼可见的鼓包与褶皱等缺陷,深度相机测量织物点云距模具点云距离小于1.0 mm的点约占97.3%。

     

    Abstract: In layup of woven composite on double curvature surface, the manual layup is of low efficiency and poor quality uniformity, while research findings on robotic layup are not accurate enough to describe manual layup skills. Therefore, a method of acquisition, representation and recurrence of woven composite layup skills is proposed based on imitation learning. Firstly, the dragging demonstration method is used to obtain trajectory information of woven layup, and supervised trajectory segmentation is carried out based on pressure threshold. After that, probabilistic models of layup trajectory are learned by Gaussian mixture model (GMM) and Gaussian mixture regression (GMR). On this basis, the normal orientation of the endeffector is calculated taking demonstration orientation and surface norm as reference. Then, another demonstration is carried out to obtain normal pressure information, and probabilistic models of layup pressure are learned though GMM/GMR. For the establishment process of contact status, a pre-control quantity exit mechanism is proposed within the framework of hybrid force/motion control to accomplish non-impact switching of the controllers during the establishment of contact status. Finally, an experiment platform is built based on UR5e robot, ATI six-dimensional force/torque sensor and ROS (robot operating system). The result shows that the proposed skills acquisition and description method can realize accurate extraction and expression of manual layup experience, and the maximum pressure error during contact status establishment is 1.8 N when using the proposed pre-control quantity exit mechanism, which is less than 2.0 N, the root mean square error of layup pressure of all 16 trajectories on the curvature surface mould. The finished layer has no visible defects such as bulge and folding. At about 97.3% of points in the woven point cloud measured by depth camera, their distances from the mould point cloud are less than 1.0 mm.

     

/

返回文章
返回