Many studies about improving the performance of biped robots, especially energy efficiency and waking speed have been done. These properties, however, are secondary to the robot ability to prevent falling down on irregular grounds. In this paper, we propose a novel method which uses an active up-and-down wobbling mass to improve walking ability and robustness of biped robot on uneven terrain. We will show improvements that the number of walk steps is increased in numerical simulations using the proposed method. We also evaluate quantitatively walking ability of biped robot on uneven terrain using a performance measure called Gait Sensitivity Norm (GSN) and ∞ norm. Furthermore, we show the effectiveness of ∞ norm measure which cannot be evaluated by GSN in a walking example.