A neural-network compensator is developed for the adaptive control of robot manipulators. The proposed compensator is implemented using the adaptive-linear-combiner algorithm with a special learning rule derived based on the Lyapunov method. Both the system stability and error convergence can be guaranteed. The resulting controller has an implementation advantage in that the adaptation part of the control structure is independent of the feedforward part of the same control algorithm and multirate sampling for the whole control system can therefore be applied. Simulation studies on a single-link manipulator show that the adaptive control system incorporated with the neurocompensator maintains a very good tracking performance even in the presence of large parameter uncertainties and external disturbance. The satisfactory control performance of this approach is also demonstrated by experimental results.