For representing the input-output behaviour of a robot manipulator by a linear time-invariant model, four direct linearization schemes are: (i) state linearization, (ii) linearization based on an identification method, (iii) linearization based on neglecting velocity-dependent and gravity terms and (iv) linearization based on neglecting the velocity-dependent term only (rate linearization). In order to make an appropriate choice of linear model for the development of real-time control, these schemes are extensively studied in this paper. It is shown that the rate linearization method leads to a satisfactory tradeoff between computation, accuracy, and stability. In the case of high velocity motions, a combination of state linearization and rate linearization is proposed. © 1993 Kluwer Academic Publishers.