Abstract: To make the virtual human＇s motion more close to real human＇s motion, a virtual human motion control system is constructed based on biomechanics. The control model is composed of three layers： the top policy control layer is responsible for monitoring the change of the environment and the motion state of the virtual human and sending the control signal to the medium layer according to the given policies; the medium spinal reflex layer can receive instructions from the top layer and the perceived information from the environment and the virtual human, and then it maps all of them to the corresponding muscle excitation signal; the bottom muscle actuated layer adopts the muscle tendon unit（MTU） model to generate the muscle force in accordance with real humans. A new evaluation strategy using gait similarity analysis is integrated to the optimization of control parameters, which makes the walking motion of the virtual human more realistically. The simulation experiments are run in the Simulink environment on Matlab. The results show that the control system is capable of generating stable and realistic walking for virtual humans and furthermore, the integration of gait similarity criterion in the process of optimization improves the realism of the walking motion.