Robot-assisted therapy has become an important technology applied in rehabilitation engineering, allowing patients with motion impairment problems to perform training programs without continuous supervision from physiotherapists. The goal of this paper is to develop a gravity balanced exoskeleton for active rehabilitation training of upper limb. The mechanical structure and kinematics of the exoskeleton are described and optimized to enable natural interaction with user and avoid singular configurations within the desired workspace. The gravity balancing of the human arm and mechanism is achieved through a hybrid strategy making use of auxiliary links and zero-free-length springs to balance the effect of gravity over the range of motion. The balance errors resulting from the variation of anthropometric parameters are analyzed and discussed. Further experiments involving trajectories tracking tasks with and without gravity balancing are conducted to evaluate the improvement of the control performance and energetic efficiency made by the developed balanced mechanism. The experimental results indicate that the proposed balance strategy can achieve a reduction of 34.56% in overall power consumption compared with the cost in unbalanced condition.