This paper presents a new variable admittance control scheme for improving physical human-robot interaction. For achieving effective assistance, an intuitive cooperation must be achieved by the human operator, in which the detection of the human inference is required. The proposed controller employs a new algorithm for inferring human intention, i.e., acceleration or deceleration, by monitoring the difference of the magnitude of the measured force during each sampling period. Then, the virtual damping is adjusted to high value for fine positioning or low value for fast motion. The proposed algorithm is validated experimentally by using a single degree of freedom device. It is compared to conventional constant-damping admittance control approach and evaluated in terms of overshoots and precise motion. It was also shown that the proposed algorithm is practically applicable to both position-based and velocity-based admittance controllers. The experimental data demonstrate that the proposed approach outperforms the conventional approach.