ニューラルネットワークを用いた反発係数の未知な対象物を連続打撃するフレキシブルリンクハンマロボット

抄録

This paper deals with a flexible link hammer which can continuously hit an object with an unknown coefficient of rebound. This hammer system makes use of only the first mode of vibration of the link for a desired hitting. The relative deflection of the flexible link hammer is expressed as a function of an angular acceleration which is used as the input of the link driver. When the hammer hits an object, the hammer has to avoid the ‘twice collision’ in one hitting cycle and flap an object with a desired hitting velocity. The expression of the relative deflection can easily derive two sets of non-linear simultaneous equations satisfying the hitting conditions. The input of the link driver are determined from the numerical solutions of the simultaneous equations. Since the equations are the function of the initial velocity, the coefficient of rebound has to be identified by using the strain of the flexible link. However, it is difficult for this numerical method to find out the optimal input pattern by on-line processing. Therefore, the multi-layered neural networks are applied for the generation of the optimal input pattern. The neural networks acquire the relationship between the initial velocity of the hammer head and the parameters of the input pattern of the hammer driver. The trained neural networks can generate the input pattern of the link driver for a measured initial velocity and are effective for the on-line processing.