This paper deals with reflexes against sudden obstacles for humanoid robots, which are combined with a proposed piecewise-linear pattern generator. The reflex action consists of modulating the motors' commands by the outputs of the force sensors located under the robot legs, and through a given primitive neural network previously proposed. Eventually, the primitive reflex is modified by the afferent signals to be more adaptable and robust against sudden obstacles. The final reflex structure enables the response of the sensory signals to be coordinated and modulated with locomotion controller's outputs to achieve an intended stabilizing behavior of the robot. The effectiveness of the proposed reflex is demonstrated by experiment using Fujitsu's humanoid robot HOAP-3. It is shown that the primitive reflex movement previously proposed becomes robust against sudden obstacle that hits the robot sole plate at different locations. The proposed reflex system, therefore, contributes towards the safer interaction of the humanoid robot with the environment.