The purpose of our study is to achieve dynamic model-based control of a nonlinear elastic-joint robot arm. We previously proposed a plug-in feedback scheme for vibration-suppression control of a serial two-link robot arm with joint elasticity due to a Harmonic-drive gear. The serial two-link arm simulates the 1st and 2nd joints of the SCARA-type robot or the 2nd and 3rd joints of the PUMA-type robot. In order to suppress the arm-tip vibration of both robot types, it is important to control the basic two-link arm. We proposed a torsion-angular-velocity feedback (TVFB) scheme, which can be plugged into existing joint servos (PI velocity controllers) using a nonlinear state-observer based on a physically parameterized dynamic model of the elastic-joint robot arm. The feedback gains of the observer are set to be identical to the PI gains tuned for the existing joint servos. Thus, the nonlinear observer, which estimates the torsion-angular velocity, is designless. This paper proposes a simple gain-scheduling scheme with a few hand-tuned state-feedback gains for improving the stability of the TVFB, taking the arm-posture and payload changes into consideration. We only have to manually tune a few state-feedback gains. Several experiments are conducted to demonstrate the vibration-suppression and fast-positioning capabilities of the improved TVFB using the elastic-joint robot arm.