This paper examines methods to estimate torque and angle in shoulder flexion from surface electromyography(sEMG) signals for intuitive and delicate control of robotic assistance device. Five muscles on the upper arm, three for shoulder flexion and two for shoulder extension, were used to offer favorable sEMG recording conditions in the estimation. The methods tested were the mean absolute value (MAV) with linear regression and the artificial neural network (ANN) method. An optimal condition was sought by varying combination of muscles used and the parameters in each method. The estimation performance was evaluated using the correlation values and normalized root mean square error values. In addition, we discussed their possible use as an estimation of motion intent of a user or as a command input in a physical human-machine interaction system.