We have applied and evaluated system identification methods using both commercial software and dedicated subspace model identification software (MOESP). Results using the different software tools have been similar (but not identical) in accuracy and predictive power, the main differences being the time required for computation and occasional failures of one algorithm in delivery of a stable model. For linear model identification all methods tested failed to provide residuals, i.e. model misfit, uncorrelated with input and without significant autocorrelation. As a result, no linear stochastic innovations model could be formulated in any satisfactory manner. However, model-order tests based on singular values suggest that a low model order be sufficient for input–output modeling to within a modeling accuracy of 2–5%. Thus, the identification of a state-space model combined with a friction model provides effective means to modeling in robotics.