Generation of human reaching movement using a recurrent neural network model

Human can spontaneously generate reasonable motion even for natural unconcerned gesture or unexperienced motion, in which we can not assume prior formation nor learning of an optimal trajectory. In this study, by contrast to the conventional trajectory planning approach, we attempt to construct a neuro-control mechanism that can spontaneously generate reaching motion without forming trajectory. The musculo-skeletal model of a human upper extremity is constructed as three rigid links with eight principal muscles. In order to represent passive joint resistance, a non-linear visco-elastic element is attached around each joint. The nervous system is modeled as a recurrent neural network which incorporates the human musculo-skeletal potential and a pseudo-potential which defines a goal position. Given a goal position, the nervous system thus generates muscular activation signals that tend to move the hand to the goal while decreasing the musculo-skeletal potential. Due to the dynamic interaction among the entire neuro-musculo-skeletal systems, the model can generate reaching movement as if hand position is attracted to the reaching goal while being naturally affected by the inherent musculo-skeletal constraints of human upper limb. Comparisons of the generated motions with measured data demonstrate that the model is capable of inducing human-like reaching motion towards a given goal position without priorly computing an optimal trajectory. The simulated result suggests that the proposed neural network model may describe a spontaneous motion generating mechanism which human may possess inherently.