This paper presents a method of utilizing estimated actual velocity of robot links for online walking pattern generation. Repetitive generation of the dynamically-feasible trajectories that start from the actual motion status of the robot is carried out, so that the balance of the actual walking will be maintained by the generation. A method of estimating the absolute velocity of the torso and feet based on the gyroscopic-sensor information is discussed. Then we present a method of designing a trajectory that satisfies estimated position and velocity as initial conditions and converges to a trajectory that realizes given walking parameters. Those 2 methods were implemented in the online walking controller of full-size humanoid HRP-2. The validity of the proposed methods is confirmed through the walking experiments of HRP-2.