Most adaptive optics systems (AO) are based on a simple control law that is unable to account for the temporal evolution of the wavefront. In this paper, a recently proposed data-driven H2-optimal control approach is demonstrated on an AO laboratory setup. The proposed control approach does not assume any form of decoupling and can therefore exploit the spatio-temporal correlation in the wavefront. The performance of the optimal control approach is compared with a conventional method. An analysis of the dominant error sources shows that the optimal control approach leads to a significant reduction in the temporal error. Since the temporal error grows with the Greenwood to sampling frequency ratio, the performance gain is especially large at large ratios.