Stroke is one of the most common diseases that lead to impairment of upper limb dexterity. Nowadays, Serious Games is a common approach to help stroke patient’s recovery. However, during motor rehabilitation protocols, it is important to detect compensatory movements, which are not currently handle in most serious games. The lack of compensatory movement detection can lead the patient to learn new and incorrect movement patterns, compromising training sessions. Thus, this paper presents a novelty technique that differentiates real upper limb functional improvements from compensatory movement patterns. To prove our theory, a system has been developed. This system consists of a highly customizable Virtual Reality serious game, with adaptable levels and tasks. Interaction with the game is done through a handle of a robotic platform. This platform has assistance feedback, which can be configured to stimulate or restrict the execution of the movements. In the game, the patient has to control a harpy to hunt and to run from predators. Inertial sensors are placed in different points of the patient paretic arm. Through these sensors, it is possible to detect a compensatory movement and to lead back the patient to correctly guide the harpy. A three-dimensional biomechanical model for musculoskeletal simulation was also developed. This model allows the simulation and analysis of muscle activity associated with movements, as well as the analysis of the kinematic synergy of the limbs, captured during the execution of training sessions. Therefore, we believe that the proposed method along with the built VR serious game system will be a useful supporting tool for helping both the patient during his training sessions and the therapist for better analysis of movements in conjunction with the definition of more specific stroke rehabilitation protocols.