The purpose of this study was to develop a new joint system based on the human shoulder mechanism. The human shoulder joint is constrained by multiple muscles and ligaments which surround the ball joint, and is driven by the balance of those forces. It is thought that the construction of the human shoulder mechanism could provide motion with three degrees of freedom and with compact size and light weight in comparison with an industrial robot arm composed of the usual pin joints. Therefore, the authors thought that the new joint mechanism for a robot arm with three degrees of freedom should be a ball joint mechanism driven by wires.In the development of the new joint mechanism, we tried to construct two types of joints. These were a five-wire model and a six-wire model. An anatomical skeletal structure has been introduced to the mechanism, especially on the muscle arrangements of the human shoulder.The movability of the new joint mechanism was evaluated by the ratio of moment arm to ball radius produced from the wires that surround the ball joint. The method of estimating moment arm was to find the slope of the wire excursion against the joint angle. Thus, to compensate some deficiency of moment arm, occurring in a specific area of flexion-extension or rotation motion, the wire position corresponding to the lack of moment arm was adjusted to an adequate position using evaluation of the moment arm. As a result of improvement processes, the effective movability was ascertained in the specified movable area for the six-wire model. Sufficient movability was not acquired in the five-wire model, however, because the movable area was considerably reduced.When the joint is driven as a robot arm, an inverse kinematics must be solved. In the present study, the problem was solved by utilizing an artificial neural network (NN), which learns data sets for arm posture and wire displacement. The additional differential outputs installed in the learned NN and consideration of a theory of virtual work balance were applied to a control system to drive the joint by a feedback control system in the range of three degrees of freedom. It was properly and precisely driven by the control system on the six-wire model. Thus, the movability and capability of the new joint system of a robot arm based on the human shoulder mechanism, as described in this paper, was demonstrated to be satisfactory.