Robotic rehabilitation has been proposed as a promising alternative in recovery after stroke, which still presents many challenges. We present here an initial approach to a progressive robot-assisted hand-motion therapy. Firstly, our system identifies finger motion patterns from electromyographic (EMG) signals of 20 control volunteers during 5 hand exercises commonly used in rehabilitation. Secondly, the system characterizes 3 muscular condition levels, using muscular contraction strength, co-activation level and muscular activation level measurements. We compared the performance of Artificial Neural Networks (ANN), Support Vector Machines (SVM), Linear Discriminant Analysis Classifier (LDA) and kNearest Neighbor (k-NN) algorithms to classify the 5 gestures and 3 levels. Thirdly, each identified gesture and level was mapped into a spatial trajectory of an exoskeleton model, using a generalization of joint trajectories from subjects and a posterior interpolation. The statistical analysis between 36 different classifier architectures showed that a SVM classifier (cubic kernel) had the best performance to identify the 15 classes (F-score of 0.8 on average). Furthermore, the average correlation between the generated spatial trajectories and the tracked hand-motion was 0.89. In the future, the trajectories controlled by EMG signals could drive the exoskeleton for rehabilitation patients.