Abstract: In this paper, according to the dynamics of running humanoid robot, a probability model of running state analysis for humanoid robot is proposed based on the feedback of virtual acceleration sensor. Inertial force affects the running state of humanoid robot during the course of running. The value of acceleration can express inertial force. So we can obtain dynamic feedback from the virtual acceleration sensor built in humanoid robot to illustrate the running state of humanoid robot, and can analyse dynamic feedback from virtual acceleration sensor by using wavelet transform and fast Fourier transform. The probability model of running state analysis for humanoid robot is formulated by energy eigenvalue abstracted in freqency field. Using Mahalanobis distance as a criteria for stable running of humanoid robot, this model can express humanoid robot running state quantitatively. Simulation is conduct for humanoid robot model built with ADAMS, and the virtual acceleration sensor is built in the center of mass for humanoid robot. The experimental results show that this model is able to describe the running of humanoid robot and express the running state of humanoid robot during the course of running including start gait and stop gait, and it can help humanoid robot adjust their gaits with the change of environment to ensure their running stability.