论文部分内容阅读
针对机器人多指手自身特点,对其所有可能具备的抓取模式进行分类,利用人工种经网络表示被抓取对象的形状、几何参数及所处姿态与抓取模式之间复杂的非线性映射关系;通过样本对人工神经网络进行训练,用训练好的人工神经网络自动生成抓取模式;讨论更为一般的力分配问题,这里手指和物体之间的接触形式可以是无摩擦的‘或接触、有摩擦的点接触及软指接触中的任何一种,并不论是否为对称抓取;为使实时抓取成为可能,考虑用BP网络和FL网络解决力分配问题,提出一种非线性规划方法,用该方法的解对BP网络和FL网络进行对比训练,发现后者能较快收敛到较小误差,仿真结果表明FL网络能实时有效规划各关节所需要的力矩。
According to the characteristics of robots’ fingerprinting, all the possible crawling patterns are classified, and the shape, geometric parameters and the nonlinear mapping between the gesture and crawling patterns Relationship; training the artificial neural network through a sample, automatically generating a fetch pattern with a trained artificial neural network; and discussing the more general problem of force distribution where the contact form between the finger and the object may be frictionless or contact , Frictional point contact and soft finger contact, whichever is symmetrical or not. In order to make real-time crawling possible, we consider the problem of force distribution with BP network and FL network and propose a nonlinear programming Method. Compared with the BP network and the FL network, the solution of this method is compared and found to converge to a smaller error quickly. The simulation results show that the FL network can effectively plan the moments required by each joint in real time.