With the gradual expansion of the development of sports, the level of sports has been rapidly improved. Athletes have to carry out high-intensity and systemic technical movements in training and competition. Some sports have the greatest burden on the shoulder joint. From the observation and investigation of the injured parts of athletes, it is found that the shoulder joint is the most common sports injury, which is the most typical sports injury. Based on the problem of insufficient strength and endurance reserve after rehabilitation of shoulder external rotator injury, it will cause muscle tension and poor extensibility. To prove the improvement effect of functional training and posture index calibration on the poor posture of the shoulder, considering the measurement of global passive torque, this paper uses a limited set of joint angles and corresponding passive torque data in the upper arm lifting trajectory to train the neural network and uses the trained network to predict the passive torque in other upper arm trajectories. The kinematics model of the shoulder joint is established, and the human-computer interaction experiment is designed on the platform of the gesture index manipulator. The passive and active torque components of the shoulder joint in the human-computer interaction process are calculated by measuring the man-machine interaction force of the subjects in the motion state, which is used as the basis for evaluating the active motion intention of the subjects. Surface electromyography (SEMG) was used to calibrate and verify the attitude index of shoulder active torque. The method proposed in this paper is helpful to achieve more efficient on-demand assisted rehabilitation training exercises, which is of great significance to improve the level of rehabilitation training.