This paper presents the modeling of the user’s dynamics synchronized with a wrist and forearm exoskeleton of 3 degrees of freedom (DoF). The biomechanical model includes components that are presented in post-stroke patients, such as viscosity and elasticity parameters. The dynamic model of the uncoupled exoskeleton is considered, so that the FE (flexion/extension), AA (adduction/abduction) and PS (pronation/supination) exercises are carried out one at a time, and not simultaneously. This work shows a comparison of the implementation of two control schemes, a PD control and an adaptive nonlinear PID-like controller. These schemes are applied to the motion control of the upper limb, which stands for wrist and forearm rehabilitation, with 3 DoF. To demonstrate the feasibility of our proposal, the user motion is included in the simulations, assuming human–robot interaction, and emulating a therapy session for patients with low and high spasticity. The results show that the adaptive nonlinear PID-like controller presents a lower mean square error than the PD control, for both low and high spasticity. The most important implications of the proposed study are, the possibility to track the smooth references, similar to a realistic therapy session and with an exoskeleton system; and the inference that adaptive controllers provide a good degree of robustness with respect to gain variations and uncertain dynamic effects. Our research is also important for designing high-end wrist and forearm exoskeletons for stroke patients, whose dynamics change over rehabilitation sessions; in this sense, typical controls, like PD, are not completely adequate to compensate these changes.