As a further step to improve the excellent tracking performance of the VLT telescopes, the intrinsic errors in the telescope drive systems are analysed. These errors fall into two categories, torque disturbances and sensor errors and they have different impact on the performance. Models for the errors are developed and algorithms for on line adaptive parameter identification are presented. The models can be used to significantly reduce the influence of the errors and also to monitor parameters like friction and unbalance. The VLT servo model is used to test and verify the models and algorithms. It follows a description of the real-time software aspects of the algorithms, which have been implemented for VxWorks-based systems. The software design allows various options for the adaptation of the process coefficients, either running permanently in background, only on demand through maintenance procedures, or fixed off-line modeling based on recorded process data. Finally, real test data are presented.