To be practical, humanoid robots must be able to manoeuvre over a variety of flat and uneven terrains, at different speeds and with varying gaits and motions. 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 to study the capabilities and limitations of various balancing algorithms when combined with a single sensor. We term this 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 (e.g. controlling multiple joints over uneven terrain), the need for more sophisticated algorithms becomes apparent.