A method for learning a torque-stroke (T-S) curve of an electric motor controlled dry clutch system is disclosed. The method includes calculating a position change value A for allowing a position change point P3 corresponding to an arbitrary torque y3 on a previous T-S curve C1 to follow-up and be moved to an expectation T-S curve C3, by a control unit, calculating a probability Pr_X3 to allow the position change value A to consider various environmental factors of a clutch within a valid range, and multiplying the probability Pr_X3 to the position change value A to calculate a final position change value A, by the control unit, and calculating a new point P3 by applying the final position change value A to the position change point P3 of the previous T-S curve C1, and generating a final T-S curve connecting the new point P3 and a touch point to learn, by the control unit.