Robust generalized likelihood ratio test based on penalization