An adaptive control scheme for mechanical manipulators is proposed. The control loop essentially consists of a network for learning the robot's inverse dynamics and on-line generating the control signal. Some simulation results are provided to evaluate the design. A supervisor is used to improve the system's performances during the adaptation transients. The supervisor exerts two supervisory actions. The first one consists basically in updating the free-design adaptive controller parameters so that the value of a quadratic loss function is maintained sufficiently small. Such a function involves past tracking errors and their predictions both on appropriate time horizons of low performances during the adaptation transients. The supervisor exerts two supervisory actions. The second supervisory action consists basically of a on-line adjustment of the sampling period within an interval centered in a nominal value of the sampling period. The sampling period is selected so that the transient of the tracking error is improved according to the simple intuitive rule of sampling faster as the tracking error changes faster.