A robot-assisted bimanual shoulder flexion rehabilitation system using surface electromyography (sEMG) for post-stroke hemiplegic patients is presented. Impedance compensation of the actuators using a disturbance observer (DOB) was applied for back-drivable operation. The sEMG signal processing was utilized to obtain a desired assistive torque. Four modes of motion (passive, mirror image, shared, and voluntary) were suggested as an appropriate training platform for the various patient statuses and levels of individual recovery. Then the performance of the impedance compensation and assistive operation of the system was verified by experiments with healthy participants. The DOB decreased resistive torque by 99% compared to the open loop performance. The shoulder torque was estimated using the sEMG and linear regression (CORR = 0.960 +/- 0.011, NRMSE = 7.31 +/- 1.32%) and an artificial neural network (ANN) (CORR = 0.986 +/- 0.005, NRMSE = 6.96 +/- 1.08%) methods for generating system input based on the user's motion intention. Every mode had less than a 6% NRMSE motion error in experiments without discomforts or resistance during shoulder flexion motion of mirror the arm.