This paper presents a new method for the position control of industrial robots with elastic joints and where the dynamic of each actuator is described by a simplified model. The inverse dynamic of the system is computed and used to compensate the nonlinear terms and decouple the system through a coordinate transformation and nonlinear feedback (exact linearization). To simplify the algorithm for the inverse dynamic and hence reduce its computation, each actuator-link pair of the robot is considered as a 2-input 2-output nonlinear system, a link subsystem and an actuator subsystem. A cascade compensation using the exact linearization is then applied to each subsystem, thereby avoiding the computation of the first and second partial derivatives of the inverse of the inertia matrix and the vector of the coriolis and centrifugal forces. This gives a formalism that is relatively simple and efficient for symbolical computation, which is very important for the maintenance of accuracy. Similarly, a cascade linear controller is constructed for each subsystem of the resulting linear decoupled 2-input 2-output system. The basis vector functions for the coordinate transformation are so chosen that only one state of the link subsystem can theoretically not be measured directly or indirectly. To estimate this state, an observer with linear error dynamic is constructed. The applicability of this observer to this general case is also proved. Simulation results using the first three links of Puma 560 are finally presented.