Recently, virtual-reality (VR) has been an emerging technology, to this regard, it is widely employed by therapists to provide rich training tasks for the purpose of motor rehabilitation in clinics. Meanwhile, along with the progress of sensing technologies as means for the interaction with virtual environment, a large amount of data, such as motor trajectory, foot pressure or electromyography, is measured via VR-based motor training tasks and is considered as important clues for functional evaluations. However, very few study thoroughly applied the sensor-based data for motor assessment, instead, evaluation scales, such as TEMPA or Fugl-Meyer, were highly relied. In this study, a VR upper-limb motor training system was proposed for stroke rehabilitation. Clinical trials with 22 stroke patients were performed to exanimate the effectiveness of the propose VR system. Moreover, a variety of motor indicators derived via motion trajectory were proposed. Further, integrating multi-model data, such as motion trajectory, task performance and evaluation scales, machine-learning method was applied to develop evidence-based assessment models in order to evaluate upper-limb motor function. The results indicated that the proposed VR system was significantly effective for motor rehabilitation. Also, a few motor indicators were found significantly different between pre and post trials and were highly correlated with the evaluation scales. Finally, with the fusion of multi-model data, the accuracy rate of machine-learning assessment model was up to 92.72% which revealed its great potential for clinical use.