A shoulder-elbow rehabilitation robot has been developed as clinical treatments to facilitate motor learning and accelerate recovery of motor functions for stroke patients. However, the connection between motor learning and muscle activation patterns for stroke patients remained unknown. This study was tried to fulfill the gap by examining the muscle coordination and motor learning strategies of normal subjects while they interacted with the rehabilitation robot. A Hill-type biomechanical model based on twelve shoulder and elbow muscles was hence constructed for the upper-limb to simulate the interaction. Two normal subjects were recruited to perform upper limb circular tracking movements, clockwise and counterclockwise, on transverse plane at shoulder level in a designed force field generated by the rehabilitation robot. From the inverse dynamics analysis, the interaction was analyzed and the patterns of muscle activation were calculated. EMG signals of eight upper limb muscles were also measured for model validation and muscle coordination observation. The principle component analysis (PCA) was performed to distinguish different groups of muscle co-activation. Results showed that the constructed biomechanical model may be used as a tool for evaluating effects of treatment and be utilized as a blueprint for the design of the training protocol for the stroke patients.