This paper studies control problems of underactuated mechanical systems with model uncertainties. The control is designed with the method of backstepping. The first-order low-pass filters are used to estimate the unknown quantities and to avoid the “explosion of terms.” A novel method is also proposed to implement the control without the knowledge of the control coefficient, which makes the whole process of backstepping control data-driven. The stability of the proposed control in the Lyapunov sense is studied. It is numerically and experimentally validated, and compared with the well-known model-based linear quadratic regulator (LQR) control. The data-driven backstepping control is found to provide comparable performances to that of the LQR control with the advantage of being model-free and robust.