This paper describes three balancing-reflex algorithms (threshold control, PID control, and hybrid control) that were implemented on a real 8 DOF small humanoid robot equipped with a two-axis accelerometer sensor. We term our approach a model-free approach, since it does not require a mathematical model of the underlying robot. Instead the controller attempts to recreate successful previous motions (so-called baseline motions). In our extensive tests, the basic threshold algorithm proves the most effective overall. All algorithms are able to balance for simple tasks, but as the balancing required becomes more complex (ie, controlling multiple joints over uneven terrain), the need for more sophisticated algorithms becomes apparent.