Even with the recent advances in the area of dynamic walking on humanoid robots there is still a significant amount of manual calibration required in practice due to the variances in the hardware. That is in order to achieve the performance needed in environments such as RoboCup. We present a LIPM-based closed-loop walk, that adapts to differences in the physical behavior of the robot by optimizing parameters of the model directly on the NAO while walking and executing other tasks. A significant amount of errors in the model predictions can be reduced without using a more complex model simply by adjusting the LIPM to fit the observed behavior. Our experiments show that the optimized model yields a more controlled, faster and even more energy-efficient walk on different NAO robots and on various surfaces without additional manual parameter tuning.