In this paper, we describe a novel and quantitative approach to assess the capability of performing training task in the third-person view virtual environment for motor rehabilitation. Our proposed approach is based on human gestures which are constructed from gesture-units according to levels of complexity. Experimented in a Cave Automatic Virtual Environment, human gestures are represented by a virtual human thus the training task of the subject is to memorize those gestures and then to reproduce them. Performance of executing this training task is measured by the similarity between the virtual human’s gesture and the subject one which is captured by an optical motion capture device. In practice, a combination of performance and the complexity of the gesture is carried out to evaluate the ability of learning the human gestures of the subject.