ANALYSIS OF GRASPING MOVEMENTS BY HUMAN HAND AND ITS APPLICATION FOR MANIPULATING HAND ROBOT

从以下三个方面研究了视觉和运动信息的整合以及人类双手抓取运动控制信号的产生机制。视觉信息对抓取成绩的影响。(1)利用虚拟三维物体图像编辑系统，研究了视觉模糊对抓取时间的影响。结果表明，模糊程度的增加延长了到达时间。(2)研究了物体状态对抓取性能的影响。以空杯、装满水的杯子、装满水和盖子的杯子为抓取对象。结果表明，物体的视觉识别概念对到达轨迹和到达时间有显著影响。改进的沙漏型神经网络模型用于综合视觉和运动信息，并在抓取过程中产生控制信号。神经网络在学习阶段自组织，并通过松弛计算生成适合抓取对象的手形。学习对象既有圆柱体、方柱、球等凸形物体，也有手柄、葫芦等凹形物体。以椭球体为测试对象，表现出较好的泛化能力。学习抓取动作动力学的神经网络模型。采用三层神经网络模型。受试者被要求牢牢抓住5个大小的物体。每次测试分别测量上臂和拇指、食指2个关节角度的4个导联肌电和握力。在学习阶段和其他几组肌电中使用它们，关节角度和抓取力作为测试数据。学习是成功的，并显示了在抓取过程中估计抓取力和手指关节扭矩的可能性。

Mechanisms of integration of visual and motor informetion and generating control signals for grasping movemetnts by human hands were investigated in the following three aspects.1. Effects of visual information on grasping performance.(1) Effects of visual blur on reaching time were investigated by using of image editting system which can show subjects virtual 3-dimensional objects. The results showed that the increase of blur extends reaching time.(2) Effects of condition of object on grasping performance were investigated. Empty cup, cup filled with water and cup with water and cap were used as grasping object. The results showed visually recognized conition of object has significant effects on reaching trajectories and time.2. Generalization ability of hour glass type neural network as a model of grasping mechanisms.Modified hour glass type neural network model were used for integrating visula and motor information and generating control signal in grasping. The neural network are selforganized in learning phase and can generate the suitable hand shapes for grasping objects by using a relaxation computation. As learning objects, we used not only convex object such as cylinder, square pillar, ball, but concave object such as handle, gourd. Rather good ability of generalization was shown by using of an ellipsoid as a test object.3. Neural network model which learns forwad dynamics of grasping movements. Three layred neural network model was used. Subjects are asked to grasp 5 sizes of object firmly. For each trial, 4 channel electromyograms from upper arm and 2 joint angles of thumb and index finger respectively are measured as well as grasping force. They are used in learning phase and other sets of EMG.joint angles and grasping force are used as test data. Learning was succesful and showed the possibility of estimating grasping force and finger joint torques during grasping.