A new golf swing robot to simulate human skill—Learning control based on direct dynamics model using recurrent ANN

The motion control skill in human hyper dynamic motions, especially in ultra high speed motions, is being investigated by the authors toward a realization of hyper dynamic manipulation by robots. As a case study, a new golf swing robot, consisting of an actuated joint and a passive joint with a mechanical stopper, has been developed by us. To improve the accuracy of the motion, learning control using ANN is introduced to shadow swing motion and the swing motion of hitting a ball, and its effectiveness has been shown by experiments. Since the model used in the method depends on the swing conditions, it is difficult to adapt the same model to all conditions. To improve the versatility of the learning control method, a recurrent ANN as a model with versatility is adopted to derive the model of direct dynamics of such strongly coupled, non-linear, non-holonomic and hybrid system. And a control method based on learning and motion generation by using the model is proposed and implemented to the golf swing robot prototype. In this paper, the method for adopting the recurrent ANN and implementing the learning control is described, and experimental results are shown.