Several practical approaches to estimate the link inertia, gravity loading, and friction coefficients of a manipulator are discussed. All of the methods have been experimentally evaluated on a six-axis robot with conventional geared electric drives. The results show that it is possible to estimate the parameters of a linearized or nonlinear rigid-body model with only a few percent error in spite of quantization effects, highly nonlinear friction and unmodeled elasticity and backslash. Keys to successful parameter estimation in this case are careful choice of test motions, low-pass filtering of the data, and reduction of the model to the truly significant parameters. It has been demonstrated that a good match between measured torques and those predicted from the model for a certain test motion is a necessary but not sufficient condition for a good model. A model that has too many parameters with respect to the test motions considered may give a good match but perform poorly in regions of the workspace where no data were collected for the estimation.>