High mechanical impedance, sensor resolution, and computing bandwidth are desirable for achieving stability in feedback control. In contrast, although the human body is physically inferior to man-made systems in terms of feedback control due to signal transmission delay, humans can achieve not only stable but also robust and adaptive control ability. For these reasons, the significance of the feedforward control of the human has been emphasized in neuroscience. In previous studies, virtual trajectory control and internal model hypotheses were used to explain the principle of human feedforward control in view of the mechanical stiffness of joints and the internal model of the brain, respectively. Inspired by these insights, in this paper, we focus on the relationship between the joint stiffness and inverse model accuracy and attempted to apply it to the field of robotics. We present a variable stiffness actuator developed using variable radius gear transmission inspired by muscle cross-bridge kinematics. The developed mechanism allows the joint stiffness to be accurately controlled without any sensor-based feedback control. Using the developed actuator, we conduct a feedforward motion generating experiment with respect to variations in the inverse dynamics model uncertainty and verified that joint stiffness can compensate for inverse model uncertainty and external disturbances. These results indicate that the developed variable stiffness actuator can be applied to the robotics field for feedforward applications and support the hypothesis that humans may utilize joint stiffness to compensate for the inverse dynamics model uncertainty.